diff --git a/crypto_kem/kyber1024/META.yml b/crypto_kem/kyber1024/META.yml
new file mode 100644
index 00000000..5055cd57
--- /dev/null
+++ b/crypto_kem/kyber1024/META.yml
@@ -0,0 +1,22 @@
+name: Kyber1024
+type: kem
+claimed-nist-level: 5
+length-public-key: 1568
+length-ciphertext: 1568
+length-secret-key: 3168
+length-shared-secret: 32
+nistkat-sha256: b4b4fc1c2cbbb182252d2822ccb8cb704bcfe876122635c5dfa48ddc09b6e73f
+principal-submitter: Peter Schwabe
+auxiliary-submitters:
+  - Roberto Avanzi
+  - Joppe Bos
+  - Léo Ducas
+  - Eike Kiltz
+  - Tancrède Lepoint
+  - Vadim Lyubashevsky
+  - John M. Schanck
+  - Gregor Seiler
+  - Damien Stehlé
+implementations:
+    - name: clean
+      version: https://github.com/pq-crystals/kyber/commit/46e283ab575ec92dfe82fb12229ae2d9d6246682
diff --git a/crypto_kem/kyber1024/clean/LICENSE b/crypto_kem/kyber1024/clean/LICENSE
new file mode 100644
index 00000000..274dc1c8
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/LICENSE
@@ -0,0 +1,15 @@
+Public Domain
+Authors: Joppe Bos,
+         Léo Ducas,
+         Eike Kiltz ,
+         Tancrède Lepoint,
+         Vadim Lyubashevsky,
+         John Schanck,
+         Peter Schwabe,
+         Gregor Seiler,
+         Damien Stehlé
+
+For Keccak and AES we are using public-domain
+code from sources and by authors listed in 
+comments on top of the respective files.
+
diff --git a/crypto_kem/kyber1024/clean/Makefile b/crypto_kem/kyber1024/clean/Makefile
new file mode 100644
index 00000000..e00741f7
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/Makefile
@@ -0,0 +1,19 @@
+# This Makefile can be used with GNU Make or BSD Make
+
+LIB=libkyber1024_clean.a
+HEADERS=api.h cbd.h indcpa.h ntt.h params.h poly.h polyvec.h reduce.h verify.h symmetric.h
+OBJECTS=cbd.o indcpa.o kem.o ntt.o poly.o polyvec.o reduce.o verify.o symmetric-fips202.o
+
+CFLAGS=-O3 -Wall -Wextra -Wpedantic -Werror -Wmissing-prototypes -std=c99 -I../../../common $(EXTRAFLAGS)
+
+all: $(LIB)
+
+%.o: %.c $(HEADERS)
+	$(CC) $(CFLAGS) -c -o $@ $<
+
+$(LIB): $(OBJECTS)
+	$(AR) -r $@ $(OBJECTS)
+
+clean:
+	$(RM) $(OBJECTS)
+	$(RM) $(LIB)
diff --git a/crypto_kem/kyber1024/clean/Makefile.Microsoft_nmake b/crypto_kem/kyber1024/clean/Makefile.Microsoft_nmake
new file mode 100644
index 00000000..ef7a5dfd
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/Makefile.Microsoft_nmake
@@ -0,0 +1,19 @@
+# This Makefile can be used with Microsoft Visual Studio's nmake using the command:
+#    nmake /f Makefile.Microsoft_nmake
+
+LIBRARY=libkyber1024_clean.lib
+OBJECTS=cbd.obj indcpa.obj kem.obj ntt.obj poly.obj polyvec.obj reduce.obj verify.obj symmetric-fips202.obj
+
+CFLAGS=/nologo /I ..\..\..\common /W4 /WX
+
+all: $(LIBRARY)
+
+# Make sure objects are recompiled if headers change.
+$(OBJECTS): *.h
+
+$(LIBRARY): $(OBJECTS)
+    LIB.EXE /NOLOGO /WX /OUT:$@ $**
+
+clean:
+    -DEL $(OBJECTS)
+    -DEL $(LIBRARY)
diff --git a/crypto_kem/kyber1024/clean/api.h b/crypto_kem/kyber1024/clean/api.h
new file mode 100644
index 00000000..17c1df98
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/api.h
@@ -0,0 +1,18 @@
+#ifndef PQCLEAN_KYBER1024_CLEAN_API_H
+#define PQCLEAN_KYBER1024_CLEAN_API_H
+
+#define PQCLEAN_KYBER1024_CLEAN_CRYPTO_SECRETKEYBYTES  3168
+#define PQCLEAN_KYBER1024_CLEAN_CRYPTO_PUBLICKEYBYTES  1568
+#define PQCLEAN_KYBER1024_CLEAN_CRYPTO_CIPHERTEXTBYTES 1568
+#define PQCLEAN_KYBER1024_CLEAN_CRYPTO_BYTES           32
+
+#define PQCLEAN_KYBER1024_CLEAN_CRYPTO_ALGNAME "Kyber1024"
+
+int PQCLEAN_KYBER1024_CLEAN_crypto_kem_keypair(unsigned char *pk, unsigned char *sk);
+
+int PQCLEAN_KYBER1024_CLEAN_crypto_kem_enc(unsigned char *ct, unsigned char *ss, const unsigned char *pk);
+
+int PQCLEAN_KYBER1024_CLEAN_crypto_kem_dec(unsigned char *ss, const unsigned char *ct, const unsigned char *sk);
+
+
+#endif
diff --git a/crypto_kem/kyber1024/clean/cbd.c b/crypto_kem/kyber1024/clean/cbd.c
new file mode 100644
index 00000000..b9eea338
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/cbd.c
@@ -0,0 +1,53 @@
+#include "cbd.h"
+
+#include "params.h"
+
+#include <stdint.h>
+
+/*************************************************
+* Name:        load32_littleendian
+*
+* Description: load bytes into a 32-bit integer
+*              in little-endian order
+*
+* Arguments:   - const unsigned char *x: pointer to input byte array
+*
+* Returns 32-bit unsigned integer loaded from x
+**************************************************/
+static uint32_t load32_littleendian(const unsigned char *x) {
+    uint32_t r;
+    r  = (uint32_t)x[0];
+    r |= (uint32_t)x[1] << 8;
+    r |= (uint32_t)x[2] << 16;
+    r |= (uint32_t)x[3] << 24;
+    return r;
+}
+
+/*************************************************
+* Name:        cbd
+*
+* Description: Given an array of uniformly random bytes, compute
+*              polynomial with coefficients distributed according to
+*              a centered binomial distribution with parameter KYBER_ETA
+*              specialized for KYBER_ETA=2
+*
+* Arguments:   - poly *r:                  pointer to output polynomial
+*              - const unsigned char *buf: pointer to input byte array
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_cbd(poly *r, const unsigned char *buf) {
+    uint32_t d, t;
+    int16_t a, b;
+    int i, j;
+
+    for (i = 0; i < KYBER_N / 8; i++) {
+        t = load32_littleendian(buf + 4 * i);
+        d  = t & 0x55555555;
+        d += (t >> 1) & 0x55555555;
+
+        for (j = 0; j < 8; j++) {
+            a = (d >>  4 * j)    & 0x3;
+            b = (d >> (4 * j + 2)) & 0x3;
+            r->coeffs[8 * i + j] = a - b;
+        }
+    }
+}
diff --git a/crypto_kem/kyber1024/clean/cbd.h b/crypto_kem/kyber1024/clean/cbd.h
new file mode 100644
index 00000000..a50cc79f
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/cbd.h
@@ -0,0 +1,8 @@
+#ifndef CBD_H
+#define CBD_H
+
+#include "poly.h"
+
+void PQCLEAN_KYBER1024_CLEAN_cbd(poly *r, const unsigned char *buf);
+
+#endif
diff --git a/crypto_kem/kyber1024/clean/indcpa.c b/crypto_kem/kyber1024/clean/indcpa.c
new file mode 100644
index 00000000..35244c1e
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/indcpa.c
@@ -0,0 +1,305 @@
+#include "indcpa.h"
+
+#include "ntt.h"
+#include "poly.h"
+#include "polyvec.h"
+#include "randombytes.h"
+#include "symmetric.h"
+
+#include <stdint.h>
+
+/*************************************************
+* Name:        pack_pk
+*
+* Description: Serialize the public key as concatenation of the
+*              serialized vector of polynomials pk
+*              and the public seed used to generate the matrix A.
+*
+* Arguments:   unsigned char *r:          pointer to the output serialized public key
+*              const poly *pk:            pointer to the input public-key polynomial
+*              const unsigned char *seed: pointer to the input public seed
+**************************************************/
+static void pack_pk(unsigned char *r, polyvec *pk, const unsigned char *seed) {
+    int i;
+    PQCLEAN_KYBER1024_CLEAN_polyvec_tobytes(r, pk);
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        r[i + KYBER_POLYVECBYTES] = seed[i];
+    }
+}
+
+/*************************************************
+* Name:        unpack_pk
+*
+* Description: De-serialize public key from a byte array;
+*              approximate inverse of pack_pk
+*
+* Arguments:   - polyvec *pk:                   pointer to output public-key vector of polynomials
+*              - unsigned char *seed:           pointer to output seed to generate matrix A
+*              - const unsigned char *packedpk: pointer to input serialized public key
+**************************************************/
+static void unpack_pk(polyvec *pk, unsigned char *seed, const unsigned char *packedpk) {
+    int i;
+    PQCLEAN_KYBER1024_CLEAN_polyvec_frombytes(pk, packedpk);
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        seed[i] = packedpk[i + KYBER_POLYVECBYTES];
+    }
+}
+
+/*************************************************
+* Name:        pack_sk
+*
+* Description: Serialize the secret key
+*
+* Arguments:   - unsigned char *r:  pointer to output serialized secret key
+*              - const polyvec *sk: pointer to input vector of polynomials (secret key)
+**************************************************/
+static void pack_sk(unsigned char *r, polyvec *sk) {
+    PQCLEAN_KYBER1024_CLEAN_polyvec_tobytes(r, sk);
+}
+
+/*************************************************
+* Name:        unpack_sk
+*
+* Description: De-serialize the secret key;
+*              inverse of pack_sk
+*
+* Arguments:   - polyvec *sk:                   pointer to output vector of polynomials (secret key)
+*              - const unsigned char *packedsk: pointer to input serialized secret key
+**************************************************/
+static void unpack_sk(polyvec *sk, const unsigned char *packedsk) {
+    PQCLEAN_KYBER1024_CLEAN_polyvec_frombytes(sk, packedsk);
+}
+
+/*************************************************
+* Name:        pack_ciphertext
+*
+* Description: Serialize the ciphertext as concatenation of the
+*              compressed and serialized vector of polynomials b
+*              and the compressed and serialized polynomial v
+*
+* Arguments:   unsigned char *r:          pointer to the output serialized ciphertext
+*              const poly *pk:            pointer to the input vector of polynomials b
+*              const unsigned char *seed: pointer to the input polynomial v
+**************************************************/
+static void pack_ciphertext(unsigned char *r, polyvec *b, poly *v) {
+    PQCLEAN_KYBER1024_CLEAN_polyvec_compress(r, b);
+    PQCLEAN_KYBER1024_CLEAN_poly_compress(r + KYBER_POLYVECCOMPRESSEDBYTES, v);
+}
+
+/*************************************************
+* Name:        unpack_ciphertext
+*
+* Description: De-serialize and decompress ciphertext from a byte array;
+*              approximate inverse of pack_ciphertext
+*
+* Arguments:   - polyvec *b:             pointer to the output vector of polynomials b
+*              - poly *v:                pointer to the output polynomial v
+*              - const unsigned char *c: pointer to the input serialized ciphertext
+**************************************************/
+static void unpack_ciphertext(polyvec *b, poly *v, const unsigned char *c) {
+    PQCLEAN_KYBER1024_CLEAN_polyvec_decompress(b, c);
+    PQCLEAN_KYBER1024_CLEAN_poly_decompress(v, c + KYBER_POLYVECCOMPRESSEDBYTES);
+}
+
+/*************************************************
+* Name:        rej_uniform
+*
+* Description: Run rejection sampling on uniform random bytes to generate
+*              uniform random integers mod q
+*
+* Arguments:   - int16_t *r:               pointer to output buffer
+*              - unsigned int len:         requested number of 16-bit integers (uniform mod q)
+*              - const unsigned char *buf: pointer to input buffer (assumed to be uniform random bytes)
+*              - unsigned int buflen:      length of input buffer in bytes
+*
+* Returns number of sampled 16-bit integers (at most len)
+**************************************************/
+static unsigned int rej_uniform(int16_t *r, unsigned int len, const unsigned char *buf, unsigned int buflen) {
+    unsigned int ctr, pos;
+    uint16_t val;
+
+    ctr = pos = 0;
+    while (ctr < len && pos + 2 <= buflen) {
+        val = buf[pos] | ((uint16_t)buf[pos + 1] << 8);
+        pos += 2;
+
+        if (val < 19 * KYBER_Q) {
+            val -= (val >> 12) * KYBER_Q; // Barrett reduction
+            r[ctr++] = (int16_t)val;
+        }
+    }
+
+    return ctr;
+}
+
+#define gen_a(A,B)  PQCLEAN_KYBER1024_CLEAN_gen_matrix(A,B,0)
+#define gen_at(A,B) PQCLEAN_KYBER1024_CLEAN_gen_matrix(A,B,1)
+
+#define GENMATRIX_MAXNBLOCKS ((530 + XOF_BLOCKBYTES) / XOF_BLOCKBYTES) /* 530 is expected number of required bytes */
+
+
+/*************************************************
+* Name:        gen_matrix
+*
+* Description: Deterministically generate matrix A (or the transpose of A)
+*              from a seed. Entries of the matrix are polynomials that look
+*              uniformly random. Performs rejection sampling on output of
+*              a XOF
+*
+* Arguments:   - polyvec *a:                pointer to ouptput matrix A
+*              - const unsigned char *seed: pointer to input seed
+*              - int transposed:            boolean deciding whether A or A^T is generated
+**************************************************/
+static void PQCLEAN_KYBER1024_CLEAN_gen_matrix(polyvec *a, const unsigned char *seed, int transposed) {
+    unsigned int ctr;
+    unsigned char i, j;
+    unsigned char buf[XOF_BLOCKBYTES * GENMATRIX_MAXNBLOCKS + 1];
+    xof_state state;
+
+    for (i = 0; i < KYBER_K; i++) {
+        for (j = 0; j < KYBER_K; j++) {
+            if (transposed) {
+                xof_absorb(&state, seed, i, j);
+            } else {
+                xof_absorb(&state, seed, j, i);
+            }
+
+            xof_squeezeblocks(buf, GENMATRIX_MAXNBLOCKS, &state);
+            ctr = rej_uniform(a[i].vec[j].coeffs, KYBER_N, buf, GENMATRIX_MAXNBLOCKS * XOF_BLOCKBYTES);
+
+            while (ctr < KYBER_N) {
+                xof_squeezeblocks(buf, 1, &state);
+                ctr += rej_uniform(a[i].vec[j].coeffs + ctr, KYBER_N - ctr, buf, XOF_BLOCKBYTES);
+            }
+        }
+    }
+}
+
+/*************************************************
+* Name:        indcpa_keypair
+*
+* Description: Generates public and private key for the CPA-secure
+*              public-key encryption scheme underlying Kyber
+*
+* Arguments:   - unsigned char *pk: pointer to output public key (of length KYBER_INDCPA_PUBLICKEYBYTES bytes)
+*              - unsigned char *sk: pointer to output private key (of length KYBER_INDCPA_SECRETKEYBYTES bytes)
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_indcpa_keypair(unsigned char *pk, unsigned char *sk) {
+    polyvec a[KYBER_K], e, pkpv, skpv;
+    unsigned char buf[2 * KYBER_SYMBYTES];
+    unsigned char *publicseed = buf;
+    unsigned char *noiseseed = buf + KYBER_SYMBYTES;
+    int i;
+    unsigned char nonce = 0;
+
+    randombytes(buf, KYBER_SYMBYTES);
+    hash_g(buf, buf, KYBER_SYMBYTES);
+
+    gen_a(a, publicseed);
+
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_getnoise(skpv.vec + i, noiseseed, nonce++);
+    }
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_getnoise(e.vec + i, noiseseed, nonce++);
+    }
+
+    PQCLEAN_KYBER1024_CLEAN_polyvec_ntt(&skpv);
+    PQCLEAN_KYBER1024_CLEAN_polyvec_ntt(&e);
+
+    // matrix-vector multiplication
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_polyvec_pointwise_acc(&pkpv.vec[i], &a[i], &skpv);
+        PQCLEAN_KYBER1024_CLEAN_poly_frommont(&pkpv.vec[i]);
+    }
+
+    PQCLEAN_KYBER1024_CLEAN_polyvec_add(&pkpv, &pkpv, &e);
+    PQCLEAN_KYBER1024_CLEAN_polyvec_reduce(&pkpv);
+
+    pack_sk(sk, &skpv);
+    pack_pk(pk, &pkpv, publicseed);
+}
+
+/*************************************************
+* Name:        indcpa_enc
+*
+* Description: Encryption function of the CPA-secure
+*              public-key encryption scheme underlying Kyber.
+*
+* Arguments:   - unsigned char *c:          pointer to output ciphertext (of length KYBER_INDCPA_BYTES bytes)
+*              - const unsigned char *m:    pointer to input message (of length KYBER_INDCPA_MSGBYTES bytes)
+*              - const unsigned char *pk:   pointer to input public key (of length KYBER_INDCPA_PUBLICKEYBYTES bytes)
+*              - const unsigned char *coin: pointer to input random coins used as seed (of length KYBER_SYMBYTES bytes)
+*                                           to deterministically generate all randomness
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_indcpa_enc(unsigned char *c,
+                                        const unsigned char *m,
+                                        const unsigned char *pk,
+                                        const unsigned char *coins) {
+    polyvec sp, pkpv, ep, at[KYBER_K], bp;
+    poly v, k, epp;
+    unsigned char seed[KYBER_SYMBYTES];
+    int i;
+    unsigned char nonce = 0;
+
+    unpack_pk(&pkpv, seed, pk);
+    PQCLEAN_KYBER1024_CLEAN_poly_frommsg(&k, m);
+    gen_at(at, seed);
+
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_getnoise(sp.vec + i, coins, nonce++);
+    }
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_getnoise(ep.vec + i, coins, nonce++);
+    }
+    PQCLEAN_KYBER1024_CLEAN_poly_getnoise(&epp, coins, nonce++);
+
+    PQCLEAN_KYBER1024_CLEAN_polyvec_ntt(&sp);
+
+    // matrix-vector multiplication
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_polyvec_pointwise_acc(&bp.vec[i], &at[i], &sp);
+    }
+
+    PQCLEAN_KYBER1024_CLEAN_polyvec_pointwise_acc(&v, &pkpv, &sp);
+
+    PQCLEAN_KYBER1024_CLEAN_polyvec_invntt(&bp);
+    PQCLEAN_KYBER1024_CLEAN_poly_invntt(&v);
+
+    PQCLEAN_KYBER1024_CLEAN_polyvec_add(&bp, &bp, &ep);
+    PQCLEAN_KYBER1024_CLEAN_poly_add(&v, &v, &epp);
+    PQCLEAN_KYBER1024_CLEAN_poly_add(&v, &v, &k);
+    PQCLEAN_KYBER1024_CLEAN_polyvec_reduce(&bp);
+    PQCLEAN_KYBER1024_CLEAN_poly_reduce(&v);
+
+    pack_ciphertext(c, &bp, &v);
+}
+
+/*************************************************
+* Name:        indcpa_dec
+*
+* Description: Decryption function of the CPA-secure
+*              public-key encryption scheme underlying Kyber.
+*
+* Arguments:   - unsigned char *m:        pointer to output decrypted message (of length KYBER_INDCPA_MSGBYTES)
+*              - const unsigned char *c:  pointer to input ciphertext (of length KYBER_INDCPA_BYTES)
+*              - const unsigned char *sk: pointer to input secret key (of length KYBER_INDCPA_SECRETKEYBYTES)
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_indcpa_dec(unsigned char *m,
+                                        const unsigned char *c,
+                                        const unsigned char *sk) {
+    polyvec bp, skpv;
+    poly v, mp;
+
+    unpack_ciphertext(&bp, &v, c);
+    unpack_sk(&skpv, sk);
+
+    PQCLEAN_KYBER1024_CLEAN_polyvec_ntt(&bp);
+    PQCLEAN_KYBER1024_CLEAN_polyvec_pointwise_acc(&mp, &skpv, &bp);
+    PQCLEAN_KYBER1024_CLEAN_poly_invntt(&mp);
+
+    PQCLEAN_KYBER1024_CLEAN_poly_sub(&mp, &v, &mp);
+    PQCLEAN_KYBER1024_CLEAN_poly_reduce(&mp);
+
+    PQCLEAN_KYBER1024_CLEAN_poly_tomsg(m, &mp);
+}
diff --git a/crypto_kem/kyber1024/clean/indcpa.h b/crypto_kem/kyber1024/clean/indcpa.h
new file mode 100644
index 00000000..13fea401
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/indcpa.h
@@ -0,0 +1,16 @@
+#ifndef INDCPA_H
+#define INDCPA_H
+
+void PQCLEAN_KYBER1024_CLEAN_indcpa_keypair(unsigned char *pk,
+        unsigned char *sk);
+
+void PQCLEAN_KYBER1024_CLEAN_indcpa_enc(unsigned char *c,
+                                        const unsigned char *m,
+                                        const unsigned char *pk,
+                                        const unsigned char *coins);
+
+void PQCLEAN_KYBER1024_CLEAN_indcpa_dec(unsigned char *m,
+                                        const unsigned char *c,
+                                        const unsigned char *sk);
+
+#endif
diff --git a/crypto_kem/kyber1024/clean/kem.c b/crypto_kem/kyber1024/clean/kem.c
new file mode 100644
index 00000000..6e50a8ac
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/kem.c
@@ -0,0 +1,100 @@
+#include "api.h"
+#include "indcpa.h"
+#include "params.h"
+#include "randombytes.h"
+#include "symmetric.h"
+#include "verify.h"
+
+#include <stdlib.h>
+
+/*************************************************
+* Name:        crypto_kem_keypair
+*
+* Description: Generates public and private key
+*              for CCA-secure Kyber key encapsulation mechanism
+*
+* Arguments:   - unsigned char *pk: pointer to output public key (an already allocated array of CRYPTO_PUBLICKEYBYTES bytes)
+*              - unsigned char *sk: pointer to output private key (an already allocated array of CRYPTO_SECRETKEYBYTES bytes)
+*
+* Returns 0 (success)
+**************************************************/
+int PQCLEAN_KYBER1024_CLEAN_crypto_kem_keypair(unsigned char *pk, unsigned char *sk) {
+    size_t i;
+    PQCLEAN_KYBER1024_CLEAN_indcpa_keypair(pk, sk);
+    for (i = 0; i < KYBER_INDCPA_PUBLICKEYBYTES; i++) {
+        sk[i + KYBER_INDCPA_SECRETKEYBYTES] = pk[i];
+    }
+    hash_h(sk + KYBER_SECRETKEYBYTES - 2 * KYBER_SYMBYTES, pk, KYBER_PUBLICKEYBYTES);
+    randombytes(sk + KYBER_SECRETKEYBYTES - KYBER_SYMBYTES, KYBER_SYMBYTES);    /* Value z for pseudo-random output on reject */
+    return 0;
+}
+
+/*************************************************
+* Name:        crypto_kem_enc
+*
+* Description: Generates cipher text and shared
+*              secret for given public key
+*
+* Arguments:   - unsigned char *ct:       pointer to output cipher text (an already allocated array of CRYPTO_CIPHERTEXTBYTES bytes)
+*              - unsigned char *ss:       pointer to output shared secret (an already allocated array of CRYPTO_BYTES bytes)
+*              - const unsigned char *pk: pointer to input public key (an already allocated array of CRYPTO_PUBLICKEYBYTES bytes)
+*
+* Returns 0 (success)
+**************************************************/
+int PQCLEAN_KYBER1024_CLEAN_crypto_kem_enc(unsigned char *ct, unsigned char *ss, const unsigned char *pk) {
+    unsigned char  kr[2 * KYBER_SYMBYTES];                                   /* Will contain key, coins */
+    unsigned char buf[2 * KYBER_SYMBYTES];
+
+    randombytes(buf, KYBER_SYMBYTES);
+    hash_h(buf, buf, KYBER_SYMBYTES);                                        /* Don't release system RNG output */
+
+    hash_h(buf + KYBER_SYMBYTES, pk, KYBER_PUBLICKEYBYTES);                  /* Multitarget countermeasure for coins + contributory KEM */
+    hash_g(kr, buf, 2 * KYBER_SYMBYTES);
+
+    PQCLEAN_KYBER1024_CLEAN_indcpa_enc(ct, buf, pk, kr + KYBER_SYMBYTES);                            /* coins are in kr+KYBER_SYMBYTES */
+
+    hash_h(kr + KYBER_SYMBYTES, ct, KYBER_CIPHERTEXTBYTES);                  /* overwrite coins in kr with H(c) */
+    kdf(ss, kr, 2 * KYBER_SYMBYTES);                                         /* hash concatenation of pre-k and H(c) to k */
+    return 0;
+}
+
+/*************************************************
+* Name:        crypto_kem_dec
+*
+* Description: Generates shared secret for given
+*              cipher text and private key
+*
+* Arguments:   - unsigned char *ss:       pointer to output shared secret (an already allocated array of CRYPTO_BYTES bytes)
+*              - const unsigned char *ct: pointer to input cipher text (an already allocated array of CRYPTO_CIPHERTEXTBYTES bytes)
+*              - const unsigned char *sk: pointer to input private key (an already allocated array of CRYPTO_SECRETKEYBYTES bytes)
+*
+* Returns 0.
+*
+* On failure, ss will contain a pseudo-random value.
+**************************************************/
+int PQCLEAN_KYBER1024_CLEAN_crypto_kem_dec(unsigned char *ss, const unsigned char *ct, const unsigned char *sk) {
+    size_t i;
+    unsigned char fail;
+    unsigned char cmp[KYBER_CIPHERTEXTBYTES];
+    unsigned char buf[2 * KYBER_SYMBYTES];
+    unsigned char kr[2 * KYBER_SYMBYTES];                                    /* Will contain key, coins */
+    const unsigned char *pk = sk + KYBER_INDCPA_SECRETKEYBYTES;
+
+    PQCLEAN_KYBER1024_CLEAN_indcpa_dec(buf, ct, sk);
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {                                   /* Multitarget countermeasure for coins + contributory KEM */
+        buf[KYBER_SYMBYTES + i] = sk[KYBER_SECRETKEYBYTES - 2 * KYBER_SYMBYTES + i];    /* Save hash by storing H(pk) in sk */
+    }
+    hash_g(kr, buf, 2 * KYBER_SYMBYTES);
+
+    PQCLEAN_KYBER1024_CLEAN_indcpa_enc(cmp, buf, pk, kr + KYBER_SYMBYTES);                           /* coins are in kr+KYBER_SYMBYTES */
+
+    fail = PQCLEAN_KYBER1024_CLEAN_verify(ct, cmp, KYBER_CIPHERTEXTBYTES);
+
+    hash_h(kr + KYBER_SYMBYTES, ct, KYBER_CIPHERTEXTBYTES);                  /* overwrite coins in kr with H(c)  */
+
+    PQCLEAN_KYBER1024_CLEAN_cmov(kr, sk + KYBER_SECRETKEYBYTES - KYBER_SYMBYTES, KYBER_SYMBYTES, fail); /* Overwrite pre-k with z on re-encryption failure */
+
+    kdf(ss, kr, 2 * KYBER_SYMBYTES);                                         /* hash concatenation of pre-k and H(c) to k */
+    return 0;
+}
diff --git a/crypto_kem/kyber1024/clean/ntt.c b/crypto_kem/kyber1024/clean/ntt.c
new file mode 100644
index 00000000..0c59d646
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/ntt.c
@@ -0,0 +1,157 @@
+#include "ntt.h"
+
+#include "params.h"
+#include "reduce.h"
+
+#include <stdint.h>
+
+/* Code to generate zetas and zetas_inv used in the number-theoretic transform:
+
+#define KYBER_ROOT_OF_UNITY 17
+
+static const uint16_t tree[128] = {
+  0, 64, 32, 96, 16, 80, 48, 112, 8, 72, 40, 104, 24, 88, 56, 120,
+  4, 68, 36, 100, 20, 84, 52, 116, 12, 76, 44, 108, 28, 92, 60, 124,
+  2, 66, 34, 98, 18, 82, 50, 114, 10, 74, 42, 106, 26, 90, 58, 122,
+  6, 70, 38, 102, 22, 86, 54, 118, 14, 78, 46, 110, 30, 94, 62, 126,
+  1, 65, 33, 97, 17, 81, 49, 113, 9, 73, 41, 105, 25, 89, 57, 121,
+  5, 69, 37, 101, 21, 85, 53, 117, 13, 77, 45, 109, 29, 93, 61, 125,
+  3, 67, 35, 99, 19, 83, 51, 115, 11, 75, 43, 107, 27, 91, 59, 123,
+  7, 71, 39, 103, 23, 87, 55, 119, 15, 79, 47, 111, 31, 95, 63, 127};
+
+
+static int16_t fqmul(int16_t a, int16_t b) {
+  return montgomery_reduce((int32_t)a*b);
+}
+
+void init_ntt() {
+  unsigned int i, j, k;
+  int16_t tmp[128];
+
+  tmp[0] = MONT;
+  for(i = 1; i < 128; ++i)
+    tmp[i] = fqmul(tmp[i-1], KYBER_ROOT_OF_UNITY*MONT % KYBER_Q);
+
+  for(i = 0; i < 128; ++i)
+    zetas[i] = tmp[tree[i]];
+
+  k = 0;
+  for(i = 64; i >= 1; i >>= 1)
+    for(j = i; j < 2*i; ++j)
+      zetas_inv[k++] = -tmp[128 - tree[j]];
+
+  zetas_inv[127] = MONT * (MONT * (KYBER_Q - 1) * ((KYBER_Q - 1)/128) % KYBER_Q) % KYBER_Q;
+}
+
+*/
+int16_t PQCLEAN_KYBER1024_CLEAN_zetas[128] = {
+    2285, 2571, 2970, 1812, 1493, 1422, 287, 202, 3158, 622, 1577, 182, 962, 2127, 1855, 1468,
+    573, 2004, 264, 383, 2500, 1458, 1727, 3199, 2648, 1017, 732, 608, 1787, 411, 3124, 1758,
+    1223, 652, 2777, 1015, 2036, 1491, 3047, 1785, 516, 3321, 3009, 2663, 1711, 2167, 126, 1469,
+    2476, 3239, 3058, 830, 107, 1908, 3082, 2378, 2931, 961, 1821, 2604, 448, 2264, 677, 2054,
+    2226, 430, 555, 843, 2078, 871, 1550, 105, 422, 587, 177, 3094, 3038, 2869, 1574, 1653,
+    3083, 778, 1159, 3182, 2552, 1483, 2727, 1119, 1739, 644, 2457, 349, 418, 329, 3173, 3254,
+    817, 1097, 603, 610, 1322, 2044, 1864, 384, 2114, 3193, 1218, 1994, 2455, 220, 2142, 1670,
+    2144, 1799, 2051, 794, 1819, 2475, 2459, 478, 3221, 3021, 996, 991, 958, 1869, 1522, 1628
+};
+
+int16_t PQCLEAN_KYBER1024_CLEAN_zetas_inv[128] = {
+    1701, 1807, 1460, 2371, 2338, 2333, 308, 108, 2851, 870, 854, 1510, 2535, 1278, 1530, 1185,
+    1659, 1187, 3109, 874, 1335, 2111, 136, 1215, 2945, 1465, 1285, 2007, 2719, 2726, 2232, 2512,
+    75, 156, 3000, 2911, 2980, 872, 2685, 1590, 2210, 602, 1846, 777, 147, 2170, 2551, 246,
+    1676, 1755, 460, 291, 235, 3152, 2742, 2907, 3224, 1779, 2458, 1251, 2486, 2774, 2899, 1103,
+    1275, 2652, 1065, 2881, 725, 1508, 2368, 398, 951, 247, 1421, 3222, 2499, 271, 90, 853,
+    1860, 3203, 1162, 1618, 666, 320, 8, 2813, 1544, 282, 1838, 1293, 2314, 552, 2677, 2106,
+    1571, 205, 2918, 1542, 2721, 2597, 2312, 681, 130, 1602, 1871, 829, 2946, 3065, 1325, 2756,
+    1861, 1474, 1202, 2367, 3147, 1752, 2707, 171, 3127, 3042, 1907, 1836, 1517, 359, 758, 1441
+};
+
+
+/*************************************************
+* Name:        fqmul
+*
+* Description: Multiplication followed by Montgomery reduction
+*
+* Arguments:   - int16_t a: first factor
+*              - int16_t b: second factor
+*
+* Returns 16-bit integer congruent to a*b*R^{-1} mod q
+**************************************************/
+static int16_t fqmul(int16_t a, int16_t b) {
+    return PQCLEAN_KYBER1024_CLEAN_montgomery_reduce((int32_t)a * b);
+}
+
+/*************************************************
+* Name:        ntt
+*
+* Description: Inplace number-theoretic transform (NTT) in Rq
+*              input is in standard order, output is in bitreversed order
+*
+* Arguments:   - int16_t *poly: pointer to input/output vector of 256 elements of Zq
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_ntt(int16_t *poly) {
+    unsigned int len, start, j, k;
+    int16_t t, zeta;
+
+    k = 1;
+    for (len = 128; len >= 2; len >>= 1) {
+        for (start = 0; start < 256; start = j + len) {
+            zeta = PQCLEAN_KYBER1024_CLEAN_zetas[k++];
+            for (j = start; j < start + len; ++j) {
+                t = fqmul(zeta, poly[j + len]);
+                poly[j + len] = poly[j] - t;
+                poly[j] = poly[j] + t;
+            }
+        }
+    }
+}
+
+/*************************************************
+* Name:        invntt
+*
+* Description: Inplace inverse number-theoretic transform in Rq
+*              input is in bitreversed order, output is in standard order
+*
+* Arguments:   - int16_t *poly: pointer to input/output vector of 256 elements of Zq
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_invntt(int16_t *poly) {
+    unsigned int start, len, j, k;
+    int16_t t, zeta;
+
+    k = 0;
+    for (len = 2; len <= 128; len <<= 1) {
+        for (start = 0; start < 256; start = j + len) {
+            zeta = PQCLEAN_KYBER1024_CLEAN_zetas_inv[k++];
+            for (j = start; j < start + len; ++j) {
+                t = poly[j];
+                poly[j] = PQCLEAN_KYBER1024_CLEAN_barrett_reduce(t + poly[j + len]);
+                poly[j + len] = t - poly[j + len];
+                poly[j + len] = fqmul(zeta, poly[j + len]);
+            }
+        }
+    }
+
+    for (j = 0; j < 256; ++j) {
+        poly[j] = fqmul(poly[j], PQCLEAN_KYBER1024_CLEAN_zetas_inv[127]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_basemul
+*
+* Description: Multiplication of polynomials in Zq[X]/((X^2-zeta))
+*              used for multiplication of elements in Rq in NTT domain
+*
+* Arguments:   - int16_t r[2]: pointer to the output polynomial
+*              - const int16_t a[2]: pointer to the first factor
+*              - const int16_t b[2]: pointer to the second factor
+*              - int16_t zeta: integer defining the reduction polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_basemul(int16_t r[2], const int16_t a[2], const int16_t b[2], int16_t zeta) {
+    r[0]  = fqmul(a[1], b[1]);
+    r[0]  = fqmul(r[0], zeta);
+    r[0] += fqmul(a[0], b[0]);
+
+    r[1]  = fqmul(a[0], b[1]);
+    r[1] += fqmul(a[1], b[0]);
+}
diff --git a/crypto_kem/kyber1024/clean/ntt.h b/crypto_kem/kyber1024/clean/ntt.h
new file mode 100644
index 00000000..0827eab6
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/ntt.h
@@ -0,0 +1,13 @@
+#ifndef NTT_H
+#define NTT_H
+
+#include <stdint.h>
+
+extern int16_t PQCLEAN_KYBER1024_CLEAN_zetas[128];
+extern int16_t PQCLEAN_KYBER1024_CLEAN_zetas_inv[128];
+
+void PQCLEAN_KYBER1024_CLEAN_ntt(int16_t *poly);
+void PQCLEAN_KYBER1024_CLEAN_invntt(int16_t *poly);
+void PQCLEAN_KYBER1024_CLEAN_basemul(int16_t r[2], const int16_t a[2], const int16_t b[2], int16_t zeta);
+
+#endif
diff --git a/crypto_kem/kyber1024/clean/params.h b/crypto_kem/kyber1024/clean/params.h
new file mode 100644
index 00000000..a3153e7e
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/params.h
@@ -0,0 +1,31 @@
+#ifndef PARAMS_H
+#define PARAMS_H
+
+#define KYBER_K 4 /* Change this for different security strengths */
+
+/* Don't change parameters below this line */
+
+#define KYBER_N 256
+#define KYBER_Q 3329
+
+#define KYBER_ETA 2
+
+#define KYBER_SYMBYTES 32   /* size in bytes of hashes, and seeds */
+#define KYBER_SSBYTES  32   /* size in bytes of shared key */
+
+#define KYBER_POLYBYTES              384
+#define KYBER_POLYVECBYTES           (KYBER_K * KYBER_POLYBYTES)
+
+#define KYBER_POLYCOMPRESSEDBYTES    160
+#define KYBER_POLYVECCOMPRESSEDBYTES (KYBER_K * 352)
+
+#define KYBER_INDCPA_MSGBYTES       KYBER_SYMBYTES
+#define KYBER_INDCPA_PUBLICKEYBYTES (KYBER_POLYVECBYTES + KYBER_SYMBYTES)
+#define KYBER_INDCPA_SECRETKEYBYTES (KYBER_POLYVECBYTES)
+#define KYBER_INDCPA_BYTES          (KYBER_POLYVECCOMPRESSEDBYTES + KYBER_POLYCOMPRESSEDBYTES)
+
+#define KYBER_PUBLICKEYBYTES  (KYBER_INDCPA_PUBLICKEYBYTES)
+#define KYBER_SECRETKEYBYTES  (KYBER_INDCPA_SECRETKEYBYTES +  KYBER_INDCPA_PUBLICKEYBYTES + 2*KYBER_SYMBYTES) /* 32 bytes of additional space to save H(pk) */
+#define KYBER_CIPHERTEXTBYTES  KYBER_INDCPA_BYTES
+
+#endif
diff --git a/crypto_kem/kyber1024/clean/poly.c b/crypto_kem/kyber1024/clean/poly.c
new file mode 100644
index 00000000..83bcb2f0
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/poly.c
@@ -0,0 +1,289 @@
+#include "poly.h"
+
+#include "cbd.h"
+#include "ntt.h"
+#include "params.h"
+#include "reduce.h"
+#include "symmetric.h"
+
+#include <stdint.h>
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_compress
+*
+* Description: Compression and subsequent serialization of a polynomial
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYCOMPRESSEDBYTES bytes)
+*              - const poly *a:    pointer to input polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_compress(unsigned char *r, poly *a) {
+    uint8_t t[8];
+    int i, j, k = 0;
+
+    PQCLEAN_KYBER1024_CLEAN_poly_csubq(a);
+
+    for (i = 0; i < KYBER_N; i += 8) {
+        for (j = 0; j < 8; j++) {
+            t[j] = ((((uint32_t)a->coeffs[i + j] << 5) + KYBER_Q / 2) / KYBER_Q) & 31;
+        }
+
+        r[k]   =  t[0]       | (t[1] << 5);
+        r[k + 1] = (t[1] >> 3) | (t[2] << 2) | (t[3] << 7);
+        r[k + 2] = (t[3] >> 1) | (t[4] << 4);
+        r[k + 3] = (t[4] >> 4) | (t[5] << 1) | (t[6] << 6);
+        r[k + 4] = (t[6] >> 2) | (t[7] << 3);
+        k += 5;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_decompress
+*
+* Description: De-serialization and subsequent decompression of a polynomial;
+*              approximate inverse of PQCLEAN_KYBER1024_CLEAN_poly_compress
+*
+* Arguments:   - poly *r:                pointer to output polynomial
+*              - const unsigned char *a: pointer to input byte array (of length KYBER_POLYCOMPRESSEDBYTES bytes)
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_decompress(poly *r, const unsigned char *a) {
+    int i;
+    for (i = 0; i < KYBER_N; i += 8) {
+        r->coeffs[i + 0] =  (((a[0] & 31) * KYBER_Q) + 16) >> 5;
+        r->coeffs[i + 1] = ((((a[0] >> 5) | ((a[1] & 3) << 3)) * KYBER_Q) + 16) >> 5;
+        r->coeffs[i + 2] = ((((a[1] >> 2) & 31) * KYBER_Q) + 16) >> 5;
+        r->coeffs[i + 3] = ((((a[1] >> 7) | ((a[2] & 15) << 1)) * KYBER_Q) + 16) >> 5;
+        r->coeffs[i + 4] = ((((a[2] >> 4) | ((a[3] &  1) << 4)) * KYBER_Q) + 16) >> 5;
+        r->coeffs[i + 5] = ((((a[3] >> 1) & 31) * KYBER_Q) + 16) >> 5;
+        r->coeffs[i + 6] = ((((a[3] >> 6) | ((a[4] &  7) << 2)) * KYBER_Q) + 16) >> 5;
+        r->coeffs[i + 7] =  (((a[4] >> 3) * KYBER_Q) + 16) >> 5;
+        a += 5;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_tobytes
+*
+* Description: Serialization of a polynomial
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYBYTES bytes)
+*              - const poly *a:    pointer to input polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_tobytes(unsigned char *r, poly *a) {
+    int i;
+    uint16_t t0, t1;
+
+    PQCLEAN_KYBER1024_CLEAN_poly_csubq(a);
+
+    for (i = 0; i < KYBER_N / 2; i++) {
+        t0 = a->coeffs[2 * i];
+        t1 = a->coeffs[2 * i + 1];
+        r[3 * i] = t0 & 0xff;
+        r[3 * i + 1] = (t0 >> 8) | ((t1 & 0xf) << 4);
+        r[3 * i + 2] = (t1 >> 4) & 0xff;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_frombytes
+*
+* Description: De-serialization of a polynomial;
+*              inverse of PQCLEAN_KYBER1024_CLEAN_poly_tobytes
+*
+* Arguments:   - poly *r:                pointer to output polynomial
+*              - const unsigned char *a: pointer to input byte array (of KYBER_POLYBYTES bytes)
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_frombytes(poly *r, const unsigned char *a) {
+    int i;
+
+    for (i = 0; i < KYBER_N / 2; i++) {
+        r->coeffs[2 * i]   = a[3 * i]        | ((uint16_t)a[3 * i + 1] & 0x0f) << 8;
+        r->coeffs[2 * i + 1] = a[3 * i + 1] >> 4 | ((uint16_t)a[3 * i + 2] & 0xff) << 4;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_getnoise
+*
+* Description: Sample a polynomial deterministically from a seed and a nonce,
+*              with output polynomial close to centered binomial distribution
+*              with parameter KYBER_ETA
+*
+* Arguments:   - poly *r:                   pointer to output polynomial
+*              - const unsigned char *seed: pointer to input seed (pointing to array of length KYBER_SYMBYTES bytes)
+*              - unsigned char nonce:       one-byte input nonce
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_getnoise(poly *r, const unsigned char *seed, unsigned char nonce) {
+    unsigned char buf[KYBER_ETA * KYBER_N / 4];
+
+    prf(buf, KYBER_ETA * KYBER_N / 4, seed, nonce);
+    PQCLEAN_KYBER1024_CLEAN_cbd(r, buf);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_ntt
+*
+* Description: Computes negacyclic number-theoretic transform (NTT) of
+*              a polynomial in place;
+*              inputs assumed to be in normal order, output in bitreversed order
+*
+* Arguments:   - uint16_t *r: pointer to in/output polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_ntt(poly *r) {
+    PQCLEAN_KYBER1024_CLEAN_ntt(r->coeffs);
+    PQCLEAN_KYBER1024_CLEAN_poly_reduce(r);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_invntt
+*
+* Description: Computes inverse of negacyclic number-theoretic transform (NTT) of
+*              a polynomial in place;
+*              inputs assumed to be in bitreversed order, output in normal order
+*
+* Arguments:   - uint16_t *a: pointer to in/output polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_invntt(poly *r) {
+    PQCLEAN_KYBER1024_CLEAN_invntt(r->coeffs);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_basemul
+*
+* Description: Multiplication of two polynomials in NTT domain
+*
+* Arguments:   - poly *r:       pointer to output polynomial
+*              - const poly *a: pointer to first input polynomial
+*              - const poly *b: pointer to second input polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_basemul(poly *r, const poly *a, const poly *b) {
+    unsigned int i;
+
+    for (i = 0; i < KYBER_N / 4; ++i) {
+        PQCLEAN_KYBER1024_CLEAN_basemul(r->coeffs + 4 * i, a->coeffs + 4 * i, b->coeffs + 4 * i, PQCLEAN_KYBER1024_CLEAN_zetas[64 + i]);
+        PQCLEAN_KYBER1024_CLEAN_basemul(r->coeffs + 4 * i + 2, a->coeffs + 4 * i + 2, b->coeffs + 4 * i + 2, - PQCLEAN_KYBER1024_CLEAN_zetas[64 + i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_frommont
+*
+* Description: Inplace conversion of all coefficients of a polynomial
+*              from Montgomery domain to normal domain
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_frommont(poly *r) {
+    int i;
+    const int16_t f = (1ULL << 32) % KYBER_Q;
+
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = PQCLEAN_KYBER1024_CLEAN_montgomery_reduce((int32_t)r->coeffs[i] * f);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_reduce
+*
+* Description: Applies Barrett reduction to all coefficients of a polynomial
+*              for details of the Barrett reduction see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_reduce(poly *r) {
+    int i;
+
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = PQCLEAN_KYBER1024_CLEAN_barrett_reduce(r->coeffs[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_csubq
+*
+* Description: Applies conditional subtraction of q to each coefficient of a polynomial
+*              for details of conditional subtraction of q see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_csubq(poly *r) {
+    int i;
+
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = PQCLEAN_KYBER1024_CLEAN_csubq(r->coeffs[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_add
+*
+* Description: Add two polynomials
+*
+* Arguments: - poly *r:       pointer to output polynomial
+*            - const poly *a: pointer to first input polynomial
+*            - const poly *b: pointer to second input polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_add(poly *r, const poly *a, const poly *b) {
+    int i;
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = a->coeffs[i] + b->coeffs[i];
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_sub
+*
+* Description: Subtract two polynomials
+*
+* Arguments: - poly *r:       pointer to output polynomial
+*            - const poly *a: pointer to first input polynomial
+*            - const poly *b: pointer to second input polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_sub(poly *r, const poly *a, const poly *b) {
+    int i;
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = a->coeffs[i] - b->coeffs[i];
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_frommsg
+*
+* Description: Convert 32-byte message to polynomial
+*
+* Arguments:   - poly *r:                  pointer to output polynomial
+*              - const unsigned char *msg: pointer to input message
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_frommsg(poly *r, const unsigned char msg[KYBER_SYMBYTES]) {
+    int i, j;
+    uint16_t mask;
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        for (j = 0; j < 8; j++) {
+            mask = -((msg[i] >> j) & 1);
+            r->coeffs[8 * i + j] = mask & ((KYBER_Q + 1) / 2);
+        }
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_poly_tomsg
+*
+* Description: Convert polynomial to 32-byte message
+*
+* Arguments:   - unsigned char *msg: pointer to output message
+*              - const poly *a:      pointer to input polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_poly_tomsg(unsigned char msg[KYBER_SYMBYTES], poly *a) {
+    uint16_t t;
+    int i, j;
+
+    PQCLEAN_KYBER1024_CLEAN_poly_csubq(a);
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        msg[i] = 0;
+        for (j = 0; j < 8; j++) {
+            t = (((a->coeffs[8 * i + j] << 1) + KYBER_Q / 2) / KYBER_Q) & 1;
+            msg[i] |= t << j;
+        }
+    }
+}
diff --git a/crypto_kem/kyber1024/clean/poly.h b/crypto_kem/kyber1024/clean/poly.h
new file mode 100644
index 00000000..2948908d
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/poly.h
@@ -0,0 +1,38 @@
+#ifndef POLY_H
+#define POLY_H
+
+#include "params.h"
+
+#include <stdint.h>
+
+/*
+ * Elements of R_q = Z_q[X]/(X^n + 1). Represents polynomial
+ * coeffs[0] + X*coeffs[1] + X^2*xoeffs[2] + ... + X^{n-1}*coeffs[n-1]
+ */
+typedef struct {
+    int16_t coeffs[KYBER_N];
+} poly;
+
+void PQCLEAN_KYBER1024_CLEAN_poly_compress(unsigned char *r, poly *a);
+void PQCLEAN_KYBER1024_CLEAN_poly_decompress(poly *r, const unsigned char *a);
+
+void PQCLEAN_KYBER1024_CLEAN_poly_tobytes(unsigned char *r, poly *a);
+void PQCLEAN_KYBER1024_CLEAN_poly_frombytes(poly *r, const unsigned char *a);
+
+void PQCLEAN_KYBER1024_CLEAN_poly_frommsg(poly *r, const unsigned char msg[KYBER_SYMBYTES]);
+void PQCLEAN_KYBER1024_CLEAN_poly_tomsg(unsigned char msg[KYBER_SYMBYTES], poly *a);
+
+void PQCLEAN_KYBER1024_CLEAN_poly_getnoise(poly *r, const unsigned char *seed, unsigned char nonce);
+
+void PQCLEAN_KYBER1024_CLEAN_poly_ntt(poly *r);
+void PQCLEAN_KYBER1024_CLEAN_poly_invntt(poly *r);
+void PQCLEAN_KYBER1024_CLEAN_poly_basemul(poly *r, const poly *a, const poly *b);
+void PQCLEAN_KYBER1024_CLEAN_poly_frommont(poly *r);
+
+void PQCLEAN_KYBER1024_CLEAN_poly_reduce(poly *r);
+void PQCLEAN_KYBER1024_CLEAN_poly_csubq(poly *r);
+
+void PQCLEAN_KYBER1024_CLEAN_poly_add(poly *r, const poly *a, const poly *b);
+void PQCLEAN_KYBER1024_CLEAN_poly_sub(poly *r, const poly *a, const poly *b);
+
+#endif
diff --git a/crypto_kem/kyber1024/clean/polyvec.c b/crypto_kem/kyber1024/clean/polyvec.c
new file mode 100644
index 00000000..101d13fb
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/polyvec.c
@@ -0,0 +1,196 @@
+#include "polyvec.h"
+
+#include "poly.h"
+
+#include <stdint.h>
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_polyvec_compress
+*
+* Description: Compress and serialize vector of polynomials
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYVECCOMPRESSEDBYTES)
+*              - const polyvec *a: pointer to input vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_polyvec_compress(unsigned char *r, polyvec *a) {
+    int i, j, k;
+
+    PQCLEAN_KYBER1024_CLEAN_polyvec_csubq(a);
+
+    uint16_t t[8];
+    for (i = 0; i < KYBER_K; i++) {
+        for (j = 0; j < KYBER_N / 8; j++) {
+            for (k = 0; k < 8; k++) {
+                t[k] = ((((uint32_t)a->vec[i].coeffs[8 * j + k] << 11) + KYBER_Q / 2) / KYBER_Q) & 0x7ff;
+            }
+
+            r[11 * j + 0] =  t[0] & 0xff;
+            r[11 * j + 1] = ((t[0] >>  8) | ((t[1] & 0x1f) << 3)) & 0xff;
+            r[11 * j + 2] = ((t[1] >>  5) | ((t[2] & 0x03) << 6)) & 0xff;
+            r[11 * j + 3] =  (t[2] >>  2) & 0xff;
+            r[11 * j + 4] = ((t[2] >> 10) | ((t[3] & 0x7f) << 1)) & 0xff;
+            r[11 * j + 5] = ((t[3] >>  7) | ((t[4] & 0x0f) << 4)) & 0xff;
+            r[11 * j + 6] = ((t[4] >>  4) | ((t[5] & 0x01) << 7)) & 0xff;
+            r[11 * j + 7] =  (t[5] >>  1) & 0xff;
+            r[11 * j + 8] = ((t[5] >>  9) | ((t[6] & 0x3f) << 2)) & 0xff;
+            r[11 * j + 9] = ((t[6] >>  6) | ((t[7] & 0x07) << 5)) & 0xff;
+            r[11 * j + 10] = (t[7] >>  3) & 0xff;
+        }
+        r += 352;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_polyvec_decompress
+*
+* Description: De-serialize and decompress vector of polynomials;
+*              approximate inverse of PQCLEAN_KYBER1024_CLEAN_polyvec_compress
+*
+* Arguments:   - polyvec *r:       pointer to output vector of polynomials
+*              - unsigned char *a: pointer to input byte array (of length KYBER_POLYVECCOMPRESSEDBYTES)
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_polyvec_decompress(polyvec *r, const unsigned char *a) {
+    int i, j;
+    for (i = 0; i < KYBER_K; i++) {
+        for (j = 0; j < KYBER_N / 8; j++) {
+            r->vec[i].coeffs[8 * j + 0] =  (((a[11 * j + 0]       | (((uint32_t)a[11 * j + 1] & 0x07) << 8)) * KYBER_Q) + 1024) >> 11;
+            r->vec[i].coeffs[8 * j + 1] = ((((a[11 * j + 1] >> 3) | (((uint32_t)a[11 * j + 2] & 0x3f) << 5)) * KYBER_Q) + 1024) >> 11;
+            r->vec[i].coeffs[8 * j + 2] = ((((a[11 * j + 2] >> 6) | (((uint32_t)a[11 * j + 3] & 0xff) << 2) | (((uint32_t)a[11 * j + 4] & 0x01) << 10)) * KYBER_Q) + 1024) >> 11;
+            r->vec[i].coeffs[8 * j + 3] = ((((a[11 * j + 4] >> 1) | (((uint32_t)a[11 * j + 5] & 0x0f) << 7)) * KYBER_Q) + 1024) >> 11;
+            r->vec[i].coeffs[8 * j + 4] = ((((a[11 * j + 5] >> 4) | (((uint32_t)a[11 * j + 6] & 0x7f) << 4)) * KYBER_Q) + 1024) >> 11;
+            r->vec[i].coeffs[8 * j + 5] = ((((a[11 * j + 6] >> 7) | (((uint32_t)a[11 * j + 7] & 0xff) << 1) | (((uint32_t)a[11 * j + 8] & 0x03) <<  9)) * KYBER_Q) + 1024) >> 11;
+            r->vec[i].coeffs[8 * j + 6] = ((((a[11 * j + 8] >> 2) | (((uint32_t)a[11 * j + 9] & 0x1f) << 6)) * KYBER_Q) + 1024) >> 11;
+            r->vec[i].coeffs[8 * j + 7] = ((((a[11 * j + 9] >> 5) | (((uint32_t)a[11 * j + 10] & 0xff) << 3)) * KYBER_Q) + 1024) >> 11;
+        }
+        a += 352;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_polyvec_tobytes
+*
+* Description: Serialize vector of polynomials
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYVECBYTES)
+*              - const polyvec *a: pointer to input vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_polyvec_tobytes(unsigned char *r, polyvec *a) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_tobytes(r + i * KYBER_POLYBYTES, &a->vec[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_polyvec_frombytes
+*
+* Description: De-serialize vector of polynomials;
+*              inverse of PQCLEAN_KYBER1024_CLEAN_polyvec_tobytes
+*
+* Arguments:   - unsigned char *r: pointer to output byte array
+*              - const polyvec *a: pointer to input vector of polynomials (of length KYBER_POLYVECBYTES)
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_polyvec_frombytes(polyvec *r, const unsigned char *a) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_frombytes(&r->vec[i], a + i * KYBER_POLYBYTES);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_polyvec_ntt
+*
+* Description: Apply forward NTT to all elements of a vector of polynomials
+*
+* Arguments:   - polyvec *r: pointer to in/output vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_polyvec_ntt(polyvec *r) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_ntt(&r->vec[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_polyvec_invntt
+*
+* Description: Apply inverse NTT to all elements of a vector of polynomials
+*
+* Arguments:   - polyvec *r: pointer to in/output vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_polyvec_invntt(polyvec *r) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_invntt(&r->vec[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_polyvec_pointwise_acc
+*
+* Description: Pointwise multiply elements of a and b and accumulate into r
+*
+* Arguments: - poly *r:          pointer to output polynomial
+*            - const polyvec *a: pointer to first input vector of polynomials
+*            - const polyvec *b: pointer to second input vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_polyvec_pointwise_acc(poly *r, const polyvec *a, const polyvec *b) {
+    int i;
+    poly t;
+
+    PQCLEAN_KYBER1024_CLEAN_poly_basemul(r, &a->vec[0], &b->vec[0]);
+    for (i = 1; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_basemul(&t, &a->vec[i], &b->vec[i]);
+        PQCLEAN_KYBER1024_CLEAN_poly_add(r, r, &t);
+    }
+
+    PQCLEAN_KYBER1024_CLEAN_poly_reduce(r);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_polyvec_reduce
+*
+* Description: Applies Barrett reduction to each coefficient
+*              of each element of a vector of polynomials
+*              for details of the Barrett reduction see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_polyvec_reduce(polyvec *r) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_reduce(&r->vec[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_polyvec_csubq
+*
+* Description: Applies conditional subtraction of q to each coefficient
+*              of each element of a vector of polynomials
+*              for details of conditional subtraction of q see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_polyvec_csubq(polyvec *r) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_csubq(&r->vec[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_polyvec_add
+*
+* Description: Add vectors of polynomials
+*
+* Arguments: - polyvec *r:       pointer to output vector of polynomials
+*            - const polyvec *a: pointer to first input vector of polynomials
+*            - const polyvec *b: pointer to second input vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_polyvec_add(polyvec *r, const polyvec *a, const polyvec *b) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER1024_CLEAN_poly_add(&r->vec[i], &a->vec[i], &b->vec[i]);
+    }
+}
diff --git a/crypto_kem/kyber1024/clean/polyvec.h b/crypto_kem/kyber1024/clean/polyvec.h
new file mode 100644
index 00000000..85d8df9c
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/polyvec.h
@@ -0,0 +1,27 @@
+#ifndef POLYVEC_H
+#define POLYVEC_H
+
+#include "params.h"
+#include "poly.h"
+
+typedef struct {
+    poly vec[KYBER_K];
+} polyvec;
+
+void PQCLEAN_KYBER1024_CLEAN_polyvec_compress(unsigned char *r, polyvec *a);
+void PQCLEAN_KYBER1024_CLEAN_polyvec_decompress(polyvec *r, const unsigned char *a);
+
+void PQCLEAN_KYBER1024_CLEAN_polyvec_tobytes(unsigned char *r, polyvec *a);
+void PQCLEAN_KYBER1024_CLEAN_polyvec_frombytes(polyvec *r, const unsigned char *a);
+
+void PQCLEAN_KYBER1024_CLEAN_polyvec_ntt(polyvec *r);
+void PQCLEAN_KYBER1024_CLEAN_polyvec_invntt(polyvec *r);
+
+void PQCLEAN_KYBER1024_CLEAN_polyvec_pointwise_acc(poly *r, const polyvec *a, const polyvec *b);
+
+void PQCLEAN_KYBER1024_CLEAN_polyvec_reduce(polyvec *r);
+void PQCLEAN_KYBER1024_CLEAN_polyvec_csubq(polyvec *r);
+
+void PQCLEAN_KYBER1024_CLEAN_polyvec_add(polyvec *r, const polyvec *a, const polyvec *b);
+
+#endif
diff --git a/crypto_kem/kyber1024/clean/reduce.c b/crypto_kem/kyber1024/clean/reduce.c
new file mode 100644
index 00000000..d704e3c9
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/reduce.c
@@ -0,0 +1,62 @@
+#include "reduce.h"
+
+#include "params.h"
+
+#include <stdint.h>
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_montgomery_reduce
+*
+* Description: Montgomery reduction; given a 32-bit integer a, computes
+*              16-bit integer congruent to a * R^-1 mod q,
+*              where R=2^16
+*
+* Arguments:   - int32_t a: input integer to be reduced; has to be in {-q2^15,...,q2^15-1}
+*
+* Returns:     integer in {-q+1,...,q-1} congruent to a * R^-1 modulo q.
+**************************************************/
+int16_t PQCLEAN_KYBER1024_CLEAN_montgomery_reduce(int32_t a) {
+    int32_t t;
+    int16_t u;
+
+    u = (int16_t)(a * QINV);
+    t = (int32_t)u * KYBER_Q;
+    t = a - t;
+    t >>= 16;
+    return (int16_t)t;
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_barrett_reduce
+*
+* Description: Barrett reduction; given a 16-bit integer a, computes
+*              16-bit integer congruent to a mod q in {0,...,q}
+*
+* Arguments:   - int16_t a: input integer to be reduced
+*
+* Returns:     integer in {0,...,q} congruent to a modulo q.
+**************************************************/
+int16_t PQCLEAN_KYBER1024_CLEAN_barrett_reduce(int16_t a) {
+    int32_t t;
+    const int32_t v = (1U << 26) / KYBER_Q + 1;
+
+    t = v * a;
+    t >>= 26;
+    t *= KYBER_Q;
+    return (int16_t)(a - t);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_csubq
+*
+* Description: Conditionallly subtract q
+*
+* Arguments:   - int16_t x: input integer
+*
+* Returns:     a - q if a >= q, else a
+**************************************************/
+int16_t PQCLEAN_KYBER1024_CLEAN_csubq(int16_t a) {
+    a -= KYBER_Q;
+    a += (a >> 15) & KYBER_Q;
+    return a;
+}
diff --git a/crypto_kem/kyber1024/clean/reduce.h b/crypto_kem/kyber1024/clean/reduce.h
new file mode 100644
index 00000000..248fd8e1
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/reduce.h
@@ -0,0 +1,15 @@
+#ifndef REDUCE_H
+#define REDUCE_H
+
+#include <stdint.h>
+
+#define MONT 2285 // 2^16 % Q
+#define QINV 62209 // q^(-1) mod 2^16
+
+int16_t PQCLEAN_KYBER1024_CLEAN_montgomery_reduce(int32_t a);
+
+int16_t PQCLEAN_KYBER1024_CLEAN_barrett_reduce(int16_t a);
+
+int16_t PQCLEAN_KYBER1024_CLEAN_csubq(int16_t a);
+
+#endif
diff --git a/crypto_kem/kyber1024/clean/symmetric-fips202.c b/crypto_kem/kyber1024/clean/symmetric-fips202.c
new file mode 100644
index 00000000..b1dbf478
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/symmetric-fips202.c
@@ -0,0 +1,64 @@
+#include "fips202.h"
+#include "symmetric.h"
+
+#include <stdlib.h>
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_kyber_shake128_absorb
+*
+* Description: Absorb step of the SHAKE128 specialized for the Kyber context.
+*
+* Arguments:   - uint64_t *s:                     pointer to (uninitialized) output Keccak state
+*              - const unsigned char *input:      pointer to KYBER_SYMBYTES input to be absorbed into s
+*              - unsigned char i                  additional byte of input
+*              - unsigned char j                  additional byte of input
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_kyber_shake128_absorb(keccak_state *s, const unsigned char *input, unsigned char x, unsigned char y) {
+    unsigned char extseed[KYBER_SYMBYTES + 2];
+    int i;
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        extseed[i] = input[i];
+    }
+    extseed[i++] = x;
+    extseed[i]   = y;
+    shake128_absorb(s->s, extseed, KYBER_SYMBYTES + 2);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_kyber_shake128_squeezeblocks
+*
+* Description: Squeeze step of SHAKE128 XOF. Squeezes full blocks of SHAKE128_RATE bytes each.
+*              Modifies the state. Can be called multiple times to keep squeezing,
+*              i.e., is incremental.
+*
+* Arguments:   - unsigned char *output:      pointer to output blocks
+*              - size_t nblocks:             number of blocks to be squeezed (written to output)
+*              - keccak_state *s:            pointer to in/output Keccak state
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_kyber_shake128_squeezeblocks(unsigned char *output, size_t nblocks, keccak_state *s) {
+    shake128_squeezeblocks(output, nblocks, s->s);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER1024_CLEAN_shake256_prf
+*
+* Description: Usage of SHAKE256 as a PRF, concatenates secret and public input
+*              and then generates outlen bytes of SHAKE256 output
+*
+* Arguments:   - unsigned char *output:      pointer to output
+*              - size_t outlen:              number of requested output bytes
+*              - const unsigned char * key:  pointer to the key (of length KYBER_SYMBYTES)
+*              - const unsigned char nonce:  single-byte nonce (public PRF input)
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_shake256_prf(unsigned char *output, size_t outlen, const unsigned char *key, unsigned char nonce) {
+    unsigned char extkey[KYBER_SYMBYTES + 1];
+    size_t i;
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        extkey[i] = key[i];
+    }
+    extkey[i] = nonce;
+
+    shake256(output, outlen, extkey, KYBER_SYMBYTES + 1);
+}
diff --git a/crypto_kem/kyber1024/clean/symmetric.h b/crypto_kem/kyber1024/clean/symmetric.h
new file mode 100644
index 00000000..c67fbd6a
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/symmetric.h
@@ -0,0 +1,28 @@
+#ifndef SYMMETRIC_H
+#define SYMMETRIC_H
+
+#include "fips202.h"
+#include "params.h"
+
+#include <stdlib.h>
+
+typedef struct {
+    uint64_t s[25];
+} keccak_state;
+
+void PQCLEAN_KYBER1024_CLEAN_kyber_shake128_absorb(keccak_state *s, const unsigned char *input, unsigned char x, unsigned char y);
+void PQCLEAN_KYBER1024_CLEAN_kyber_shake128_squeezeblocks(unsigned char *output, size_t nblocks, keccak_state *s);
+void PQCLEAN_KYBER1024_CLEAN_shake256_prf(unsigned char *output, size_t outlen, const unsigned char *key, unsigned char nonce);
+
+#define hash_h(OUT, IN, INBYTES) sha3_256(OUT, IN, INBYTES)
+#define hash_g(OUT, IN, INBYTES) sha3_512(OUT, IN, INBYTES)
+#define xof_absorb(STATE, IN, X, Y) PQCLEAN_KYBER1024_CLEAN_kyber_shake128_absorb(STATE, IN, X, Y)
+#define xof_squeezeblocks(OUT, OUTBLOCKS, STATE) PQCLEAN_KYBER1024_CLEAN_kyber_shake128_squeezeblocks(OUT, OUTBLOCKS, STATE)
+#define prf(OUT, OUTBYTES, KEY, NONCE) PQCLEAN_KYBER1024_CLEAN_shake256_prf(OUT, OUTBYTES, KEY, NONCE)
+#define kdf(OUT, IN, INBYTES) shake256(OUT, KYBER_SSBYTES, IN, INBYTES)
+
+#define XOF_BLOCKBYTES 168
+
+typedef keccak_state xof_state;
+
+#endif /* SYMMETRIC_H */
diff --git a/crypto_kem/kyber1024/clean/verify.c b/crypto_kem/kyber1024/clean/verify.c
new file mode 100644
index 00000000..1b580c8f
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/verify.c
@@ -0,0 +1,51 @@
+#include "verify.h"
+
+#include <stdint.h>
+#include <stdlib.h>
+
+/*************************************************
+* Name:        verify
+*
+* Description: Compare two arrays for equality in constant time.
+*
+* Arguments:   const unsigned char *a: pointer to first byte array
+*              const unsigned char *b: pointer to second byte array
+*              size_t len:             length of the byte arrays
+*
+* Returns 0 if the byte arrays are equal, 1 otherwise
+**************************************************/
+unsigned char PQCLEAN_KYBER1024_CLEAN_verify(const unsigned char *a, const unsigned char *b, size_t len) {
+    uint64_t r;
+    size_t i;
+
+    r = 0;
+    for (i = 0; i < len; i++) {
+        r |= a[i] ^ b[i];
+    }
+
+    r = (~r + 1); // Two's complement
+    r >>= 63;
+    return (unsigned char)r;
+}
+
+/*************************************************
+* Name:        cmov
+*
+* Description: Copy len bytes from x to r if b is 1;
+*              don't modify x if b is 0. Requires b to be in {0,1};
+*              assumes two's complement representation of negative integers.
+*              Runs in constant time.
+*
+* Arguments:   unsigned char *r:       pointer to output byte array
+*              const unsigned char *x: pointer to input byte array
+*              size_t len:             Amount of bytes to be copied
+*              unsigned char b:        Condition bit; has to be in {0,1}
+**************************************************/
+void PQCLEAN_KYBER1024_CLEAN_cmov(unsigned char *r, const unsigned char *x, size_t len, unsigned char b) {
+    size_t i;
+
+    b = -b;
+    for (i = 0; i < len; i++) {
+        r[i] ^= b & (x[i] ^ r[i]);
+    }
+}
diff --git a/crypto_kem/kyber1024/clean/verify.h b/crypto_kem/kyber1024/clean/verify.h
new file mode 100644
index 00000000..d55b09ac
--- /dev/null
+++ b/crypto_kem/kyber1024/clean/verify.h
@@ -0,0 +1,10 @@
+#ifndef VERIFY_H
+#define VERIFY_H
+
+#include <stdio.h>
+
+unsigned char PQCLEAN_KYBER1024_CLEAN_verify(const unsigned char *a, const unsigned char *b, size_t len);
+
+void PQCLEAN_KYBER1024_CLEAN_cmov(unsigned char *r, const unsigned char *x, size_t len, unsigned char b);
+
+#endif
diff --git a/crypto_kem/kyber512/META.yml b/crypto_kem/kyber512/META.yml
new file mode 100644
index 00000000..47148557
--- /dev/null
+++ b/crypto_kem/kyber512/META.yml
@@ -0,0 +1,22 @@
+name: Kyber512
+type: kem
+claimed-nist-level: 1
+length-public-key: 800
+length-ciphertext: 736
+length-secret-key: 1632
+length-shared-secret: 32
+nistkat-sha256: bdd9b46001de4595a4f185aec8d5d04d217705e65e10711c99fa3f0ac3d61c21
+principal-submitter: Peter Schwabe
+auxiliary-submitters:
+  - Roberto Avanzi
+  - Joppe Bos
+  - Léo Ducas
+  - Eike Kiltz
+  - Tancrède Lepoint
+  - Vadim Lyubashevsky
+  - John M. Schanck
+  - Gregor Seiler
+  - Damien Stehlé
+implementations:
+    - name: clean
+      version: https://github.com/pq-crystals/kyber/commit/46e283ab575ec92dfe82fb12229ae2d9d6246682
diff --git a/crypto_kem/kyber512/clean/LICENSE b/crypto_kem/kyber512/clean/LICENSE
new file mode 100644
index 00000000..274dc1c8
--- /dev/null
+++ b/crypto_kem/kyber512/clean/LICENSE
@@ -0,0 +1,15 @@
+Public Domain
+Authors: Joppe Bos,
+         Léo Ducas,
+         Eike Kiltz ,
+         Tancrède Lepoint,
+         Vadim Lyubashevsky,
+         John Schanck,
+         Peter Schwabe,
+         Gregor Seiler,
+         Damien Stehlé
+
+For Keccak and AES we are using public-domain
+code from sources and by authors listed in 
+comments on top of the respective files.
+
diff --git a/crypto_kem/kyber512/clean/Makefile b/crypto_kem/kyber512/clean/Makefile
new file mode 100644
index 00000000..84dc881b
--- /dev/null
+++ b/crypto_kem/kyber512/clean/Makefile
@@ -0,0 +1,19 @@
+# This Makefile can be used with GNU Make or BSD Make
+
+LIB=libkyber512_clean.a
+HEADERS=api.h cbd.h indcpa.h ntt.h params.h poly.h polyvec.h reduce.h verify.h symmetric.h
+OBJECTS=cbd.o indcpa.o kem.o ntt.o poly.o polyvec.o reduce.o verify.o symmetric-fips202.o
+
+CFLAGS=-Wall -O3 -Wmissing-prototypes -Wextra -Wpedantic -Werror -std=c99 -I../../../common $(EXTRAFLAGS)
+
+all: $(LIB)
+
+%.o: %.c $(HEADERS)
+	$(CC) $(CFLAGS) -c -o $@ $<
+
+$(LIB): $(OBJECTS)
+	$(AR) -r $@ $(OBJECTS)
+
+clean:
+	$(RM) $(OBJECTS)
+	$(RM) $(LIB)
diff --git a/crypto_kem/kyber512/clean/Makefile.Microsoft_nmake b/crypto_kem/kyber512/clean/Makefile.Microsoft_nmake
new file mode 100644
index 00000000..cead5b32
--- /dev/null
+++ b/crypto_kem/kyber512/clean/Makefile.Microsoft_nmake
@@ -0,0 +1,19 @@
+# This Makefile can be used with Microsoft Visual Studio's nmake using the command:
+#    nmake /f Makefile.Microsoft_nmake
+
+LIBRARY=libkyber512_clean.lib
+OBJECTS=cbd.obj indcpa.obj kem.obj ntt.obj poly.obj polyvec.obj reduce.obj verify.obj symmetric-fips202.obj
+
+CFLAGS=/nologo /I ..\..\..\common /W4 /WX
+
+all: $(LIBRARY)
+
+# Make sure objects are recompiled if headers change.
+$(OBJECTS): *.h
+
+$(LIBRARY): $(OBJECTS)
+    LIB.EXE /NOLOGO /WX /OUT:$@ $**
+
+clean:
+    -DEL $(OBJECTS)
+    -DEL $(LIBRARY)
diff --git a/crypto_kem/kyber512/clean/api.h b/crypto_kem/kyber512/clean/api.h
new file mode 100644
index 00000000..57c8f3bd
--- /dev/null
+++ b/crypto_kem/kyber512/clean/api.h
@@ -0,0 +1,18 @@
+#ifndef PQCLEAN_KYBER512_CLEAN_API_H
+#define PQCLEAN_KYBER512_CLEAN_API_H
+
+#define PQCLEAN_KYBER512_CLEAN_CRYPTO_SECRETKEYBYTES  1632
+#define PQCLEAN_KYBER512_CLEAN_CRYPTO_PUBLICKEYBYTES  800
+#define PQCLEAN_KYBER512_CLEAN_CRYPTO_CIPHERTEXTBYTES 736
+#define PQCLEAN_KYBER512_CLEAN_CRYPTO_BYTES           32
+
+#define PQCLEAN_KYBER512_CLEAN_CRYPTO_ALGNAME "Kyber512"
+
+int PQCLEAN_KYBER512_CLEAN_crypto_kem_keypair(unsigned char *pk, unsigned char *sk);
+
+int PQCLEAN_KYBER512_CLEAN_crypto_kem_enc(unsigned char *ct, unsigned char *ss, const unsigned char *pk);
+
+int PQCLEAN_KYBER512_CLEAN_crypto_kem_dec(unsigned char *ss, const unsigned char *ct, const unsigned char *sk);
+
+
+#endif
diff --git a/crypto_kem/kyber512/clean/cbd.c b/crypto_kem/kyber512/clean/cbd.c
new file mode 100644
index 00000000..b1567c5b
--- /dev/null
+++ b/crypto_kem/kyber512/clean/cbd.c
@@ -0,0 +1,53 @@
+#include "cbd.h"
+
+#include "params.h"
+
+#include <stdint.h>
+
+/*************************************************
+* Name:        load32_littleendian
+*
+* Description: load bytes into a 32-bit integer
+*              in little-endian order
+*
+* Arguments:   - const unsigned char *x: pointer to input byte array
+*
+* Returns 32-bit unsigned integer loaded from x
+**************************************************/
+static uint32_t load32_littleendian(const unsigned char *x) {
+    uint32_t r;
+    r  = (uint32_t)x[0];
+    r |= (uint32_t)x[1] << 8;
+    r |= (uint32_t)x[2] << 16;
+    r |= (uint32_t)x[3] << 24;
+    return r;
+}
+
+/*************************************************
+* Name:        cbd
+*
+* Description: Given an array of uniformly random bytes, compute
+*              polynomial with coefficients distributed according to
+*              a centered binomial distribution with parameter KYBER_ETA
+*              specialized for KYBER_ETA=2
+*
+* Arguments:   - poly *r:                  pointer to output polynomial
+*              - const unsigned char *buf: pointer to input byte array
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_cbd(poly *r, const unsigned char *buf) {
+    uint32_t d, t;
+    int16_t a, b;
+    int i, j;
+
+    for (i = 0; i < KYBER_N / 8; i++) {
+        t = load32_littleendian(buf + 4 * i);
+        d  = t & 0x55555555;
+        d += (t >> 1) & 0x55555555;
+
+        for (j = 0; j < 8; j++) {
+            a = (d >>  4 * j)    & 0x3;
+            b = (d >> (4 * j + 2)) & 0x3;
+            r->coeffs[8 * i + j] = a - b;
+        }
+    }
+}
diff --git a/crypto_kem/kyber512/clean/cbd.h b/crypto_kem/kyber512/clean/cbd.h
new file mode 100644
index 00000000..2ed43edc
--- /dev/null
+++ b/crypto_kem/kyber512/clean/cbd.h
@@ -0,0 +1,8 @@
+#ifndef CBD_H
+#define CBD_H
+
+#include "poly.h"
+
+void PQCLEAN_KYBER512_CLEAN_cbd(poly *r, const unsigned char *buf);
+
+#endif
diff --git a/crypto_kem/kyber512/clean/indcpa.c b/crypto_kem/kyber512/clean/indcpa.c
new file mode 100644
index 00000000..7bb6867b
--- /dev/null
+++ b/crypto_kem/kyber512/clean/indcpa.c
@@ -0,0 +1,305 @@
+#include "indcpa.h"
+
+#include "ntt.h"
+#include "poly.h"
+#include "polyvec.h"
+#include "randombytes.h"
+#include "symmetric.h"
+
+#include <stdint.h>
+
+/*************************************************
+* Name:        pack_pk
+*
+* Description: Serialize the public key as concatenation of the
+*              serialized vector of polynomials pk
+*              and the public seed used to generate the matrix A.
+*
+* Arguments:   unsigned char *r:          pointer to the output serialized public key
+*              const poly *pk:            pointer to the input public-key polynomial
+*              const unsigned char *seed: pointer to the input public seed
+**************************************************/
+static void pack_pk(unsigned char *r, polyvec *pk, const unsigned char *seed) {
+    int i;
+    PQCLEAN_KYBER512_CLEAN_polyvec_tobytes(r, pk);
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        r[i + KYBER_POLYVECBYTES] = seed[i];
+    }
+}
+
+/*************************************************
+* Name:        unpack_pk
+*
+* Description: De-serialize public key from a byte array;
+*              approximate inverse of pack_pk
+*
+* Arguments:   - polyvec *pk:                   pointer to output public-key vector of polynomials
+*              - unsigned char *seed:           pointer to output seed to generate matrix A
+*              - const unsigned char *packedpk: pointer to input serialized public key
+**************************************************/
+static void unpack_pk(polyvec *pk, unsigned char *seed, const unsigned char *packedpk) {
+    int i;
+    PQCLEAN_KYBER512_CLEAN_polyvec_frombytes(pk, packedpk);
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        seed[i] = packedpk[i + KYBER_POLYVECBYTES];
+    }
+}
+
+/*************************************************
+* Name:        pack_sk
+*
+* Description: Serialize the secret key
+*
+* Arguments:   - unsigned char *r:  pointer to output serialized secret key
+*              - const polyvec *sk: pointer to input vector of polynomials (secret key)
+**************************************************/
+static void pack_sk(unsigned char *r, polyvec *sk) {
+    PQCLEAN_KYBER512_CLEAN_polyvec_tobytes(r, sk);
+}
+
+/*************************************************
+* Name:        unpack_sk
+*
+* Description: De-serialize the secret key;
+*              inverse of pack_sk
+*
+* Arguments:   - polyvec *sk:                   pointer to output vector of polynomials (secret key)
+*              - const unsigned char *packedsk: pointer to input serialized secret key
+**************************************************/
+static void unpack_sk(polyvec *sk, const unsigned char *packedsk) {
+    PQCLEAN_KYBER512_CLEAN_polyvec_frombytes(sk, packedsk);
+}
+
+/*************************************************
+* Name:        pack_ciphertext
+*
+* Description: Serialize the ciphertext as concatenation of the
+*              compressed and serialized vector of polynomials b
+*              and the compressed and serialized polynomial v
+*
+* Arguments:   unsigned char *r:          pointer to the output serialized ciphertext
+*              const poly *pk:            pointer to the input vector of polynomials b
+*              const unsigned char *seed: pointer to the input polynomial v
+**************************************************/
+static void pack_ciphertext(unsigned char *r, polyvec *b, poly *v) {
+    PQCLEAN_KYBER512_CLEAN_polyvec_compress(r, b);
+    PQCLEAN_KYBER512_CLEAN_poly_compress(r + KYBER_POLYVECCOMPRESSEDBYTES, v);
+}
+
+/*************************************************
+* Name:        unpack_ciphertext
+*
+* Description: De-serialize and decompress ciphertext from a byte array;
+*              approximate inverse of pack_ciphertext
+*
+* Arguments:   - polyvec *b:             pointer to the output vector of polynomials b
+*              - poly *v:                pointer to the output polynomial v
+*              - const unsigned char *c: pointer to the input serialized ciphertext
+**************************************************/
+static void unpack_ciphertext(polyvec *b, poly *v, const unsigned char *c) {
+    PQCLEAN_KYBER512_CLEAN_polyvec_decompress(b, c);
+    PQCLEAN_KYBER512_CLEAN_poly_decompress(v, c + KYBER_POLYVECCOMPRESSEDBYTES);
+}
+
+/*************************************************
+* Name:        rej_uniform
+*
+* Description: Run rejection sampling on uniform random bytes to generate
+*              uniform random integers mod q
+*
+* Arguments:   - int16_t *r:               pointer to output buffer
+*              - unsigned int len:         requested number of 16-bit integers (uniform mod q)
+*              - const unsigned char *buf: pointer to input buffer (assumed to be uniform random bytes)
+*              - unsigned int buflen:      length of input buffer in bytes
+*
+* Returns number of sampled 16-bit integers (at most len)
+**************************************************/
+static unsigned int rej_uniform(int16_t *r, unsigned int len, const unsigned char *buf, unsigned int buflen) {
+    unsigned int ctr, pos;
+    uint16_t val;
+
+    ctr = pos = 0;
+    while (ctr < len && pos + 2 <= buflen) {
+        val = buf[pos] | ((uint16_t)buf[pos + 1] << 8);
+        pos += 2;
+
+        if (val < 19 * KYBER_Q) {
+            val -= (val >> 12) * KYBER_Q; // Barrett reduction
+            r[ctr++] = (int16_t)val;
+        }
+    }
+
+    return ctr;
+}
+
+#define gen_a(A,B)  PQCLEAN_KYBER512_CLEAN_gen_matrix(A,B,0)
+#define gen_at(A,B) PQCLEAN_KYBER512_CLEAN_gen_matrix(A,B,1)
+
+#define GENMATRIX_MAXNBLOCKS ((530 + XOF_BLOCKBYTES) / XOF_BLOCKBYTES) /* 530 is expected number of required bytes */
+
+
+/*************************************************
+* Name:        gen_matrix
+*
+* Description: Deterministically generate matrix A (or the transpose of A)
+*              from a seed. Entries of the matrix are polynomials that look
+*              uniformly random. Performs rejection sampling on output of
+*              a XOF
+*
+* Arguments:   - polyvec *a:                pointer to ouptput matrix A
+*              - const unsigned char *seed: pointer to input seed
+*              - int transposed:            boolean deciding whether A or A^T is generated
+**************************************************/
+static void PQCLEAN_KYBER512_CLEAN_gen_matrix(polyvec *a, const unsigned char *seed, int transposed) {
+    unsigned int ctr;
+    unsigned char i, j;
+    unsigned char buf[XOF_BLOCKBYTES * GENMATRIX_MAXNBLOCKS + 1];
+    xof_state state;
+
+    for (i = 0; i < KYBER_K; i++) {
+        for (j = 0; j < KYBER_K; j++) {
+            if (transposed) {
+                xof_absorb(&state, seed, i, j);
+            } else {
+                xof_absorb(&state, seed, j, i);
+            }
+
+            xof_squeezeblocks(buf, GENMATRIX_MAXNBLOCKS, &state);
+            ctr = rej_uniform(a[i].vec[j].coeffs, KYBER_N, buf, GENMATRIX_MAXNBLOCKS * XOF_BLOCKBYTES);
+
+            while (ctr < KYBER_N) {
+                xof_squeezeblocks(buf, 1, &state);
+                ctr += rej_uniform(a[i].vec[j].coeffs + ctr, KYBER_N - ctr, buf, XOF_BLOCKBYTES);
+            }
+        }
+    }
+}
+
+/*************************************************
+* Name:        indcpa_keypair
+*
+* Description: Generates public and private key for the CPA-secure
+*              public-key encryption scheme underlying Kyber
+*
+* Arguments:   - unsigned char *pk: pointer to output public key (of length KYBER_INDCPA_PUBLICKEYBYTES bytes)
+*              - unsigned char *sk: pointer to output private key (of length KYBER_INDCPA_SECRETKEYBYTES bytes)
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_indcpa_keypair(unsigned char *pk, unsigned char *sk) {
+    polyvec a[KYBER_K], e, pkpv, skpv;
+    unsigned char buf[2 * KYBER_SYMBYTES];
+    unsigned char *publicseed = buf;
+    unsigned char *noiseseed = buf + KYBER_SYMBYTES;
+    int i;
+    unsigned char nonce = 0;
+
+    randombytes(buf, KYBER_SYMBYTES);
+    hash_g(buf, buf, KYBER_SYMBYTES);
+
+    gen_a(a, publicseed);
+
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_getnoise(skpv.vec + i, noiseseed, nonce++);
+    }
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_getnoise(e.vec + i, noiseseed, nonce++);
+    }
+
+    PQCLEAN_KYBER512_CLEAN_polyvec_ntt(&skpv);
+    PQCLEAN_KYBER512_CLEAN_polyvec_ntt(&e);
+
+    // matrix-vector multiplication
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_polyvec_pointwise_acc(&pkpv.vec[i], &a[i], &skpv);
+        PQCLEAN_KYBER512_CLEAN_poly_frommont(&pkpv.vec[i]);
+    }
+
+    PQCLEAN_KYBER512_CLEAN_polyvec_add(&pkpv, &pkpv, &e);
+    PQCLEAN_KYBER512_CLEAN_polyvec_reduce(&pkpv);
+
+    pack_sk(sk, &skpv);
+    pack_pk(pk, &pkpv, publicseed);
+}
+
+/*************************************************
+* Name:        indcpa_enc
+*
+* Description: Encryption function of the CPA-secure
+*              public-key encryption scheme underlying Kyber.
+*
+* Arguments:   - unsigned char *c:          pointer to output ciphertext (of length KYBER_INDCPA_BYTES bytes)
+*              - const unsigned char *m:    pointer to input message (of length KYBER_INDCPA_MSGBYTES bytes)
+*              - const unsigned char *pk:   pointer to input public key (of length KYBER_INDCPA_PUBLICKEYBYTES bytes)
+*              - const unsigned char *coin: pointer to input random coins used as seed (of length KYBER_SYMBYTES bytes)
+*                                           to deterministically generate all randomness
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_indcpa_enc(unsigned char *c,
+                                       const unsigned char *m,
+                                       const unsigned char *pk,
+                                       const unsigned char *coins) {
+    polyvec sp, pkpv, ep, at[KYBER_K], bp;
+    poly v, k, epp;
+    unsigned char seed[KYBER_SYMBYTES];
+    int i;
+    unsigned char nonce = 0;
+
+    unpack_pk(&pkpv, seed, pk);
+    PQCLEAN_KYBER512_CLEAN_poly_frommsg(&k, m);
+    gen_at(at, seed);
+
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_getnoise(sp.vec + i, coins, nonce++);
+    }
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_getnoise(ep.vec + i, coins, nonce++);
+    }
+    PQCLEAN_KYBER512_CLEAN_poly_getnoise(&epp, coins, nonce++);
+
+    PQCLEAN_KYBER512_CLEAN_polyvec_ntt(&sp);
+
+    // matrix-vector multiplication
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_polyvec_pointwise_acc(&bp.vec[i], &at[i], &sp);
+    }
+
+    PQCLEAN_KYBER512_CLEAN_polyvec_pointwise_acc(&v, &pkpv, &sp);
+
+    PQCLEAN_KYBER512_CLEAN_polyvec_invntt(&bp);
+    PQCLEAN_KYBER512_CLEAN_poly_invntt(&v);
+
+    PQCLEAN_KYBER512_CLEAN_polyvec_add(&bp, &bp, &ep);
+    PQCLEAN_KYBER512_CLEAN_poly_add(&v, &v, &epp);
+    PQCLEAN_KYBER512_CLEAN_poly_add(&v, &v, &k);
+    PQCLEAN_KYBER512_CLEAN_polyvec_reduce(&bp);
+    PQCLEAN_KYBER512_CLEAN_poly_reduce(&v);
+
+    pack_ciphertext(c, &bp, &v);
+}
+
+/*************************************************
+* Name:        indcpa_dec
+*
+* Description: Decryption function of the CPA-secure
+*              public-key encryption scheme underlying Kyber.
+*
+* Arguments:   - unsigned char *m:        pointer to output decrypted message (of length KYBER_INDCPA_MSGBYTES)
+*              - const unsigned char *c:  pointer to input ciphertext (of length KYBER_INDCPA_BYTES)
+*              - const unsigned char *sk: pointer to input secret key (of length KYBER_INDCPA_SECRETKEYBYTES)
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_indcpa_dec(unsigned char *m,
+                                       const unsigned char *c,
+                                       const unsigned char *sk) {
+    polyvec bp, skpv;
+    poly v, mp;
+
+    unpack_ciphertext(&bp, &v, c);
+    unpack_sk(&skpv, sk);
+
+    PQCLEAN_KYBER512_CLEAN_polyvec_ntt(&bp);
+    PQCLEAN_KYBER512_CLEAN_polyvec_pointwise_acc(&mp, &skpv, &bp);
+    PQCLEAN_KYBER512_CLEAN_poly_invntt(&mp);
+
+    PQCLEAN_KYBER512_CLEAN_poly_sub(&mp, &v, &mp);
+    PQCLEAN_KYBER512_CLEAN_poly_reduce(&mp);
+
+    PQCLEAN_KYBER512_CLEAN_poly_tomsg(m, &mp);
+}
diff --git a/crypto_kem/kyber512/clean/indcpa.h b/crypto_kem/kyber512/clean/indcpa.h
new file mode 100644
index 00000000..2d77ff22
--- /dev/null
+++ b/crypto_kem/kyber512/clean/indcpa.h
@@ -0,0 +1,16 @@
+#ifndef INDCPA_H
+#define INDCPA_H
+
+void PQCLEAN_KYBER512_CLEAN_indcpa_keypair(unsigned char *pk,
+        unsigned char *sk);
+
+void PQCLEAN_KYBER512_CLEAN_indcpa_enc(unsigned char *c,
+                                       const unsigned char *m,
+                                       const unsigned char *pk,
+                                       const unsigned char *coins);
+
+void PQCLEAN_KYBER512_CLEAN_indcpa_dec(unsigned char *m,
+                                       const unsigned char *c,
+                                       const unsigned char *sk);
+
+#endif
diff --git a/crypto_kem/kyber512/clean/kem.c b/crypto_kem/kyber512/clean/kem.c
new file mode 100644
index 00000000..175a9a71
--- /dev/null
+++ b/crypto_kem/kyber512/clean/kem.c
@@ -0,0 +1,100 @@
+#include "api.h"
+#include "indcpa.h"
+#include "params.h"
+#include "randombytes.h"
+#include "symmetric.h"
+#include "verify.h"
+
+#include <stdlib.h>
+
+/*************************************************
+* Name:        crypto_kem_keypair
+*
+* Description: Generates public and private key
+*              for CCA-secure Kyber key encapsulation mechanism
+*
+* Arguments:   - unsigned char *pk: pointer to output public key (an already allocated array of CRYPTO_PUBLICKEYBYTES bytes)
+*              - unsigned char *sk: pointer to output private key (an already allocated array of CRYPTO_SECRETKEYBYTES bytes)
+*
+* Returns 0 (success)
+**************************************************/
+int PQCLEAN_KYBER512_CLEAN_crypto_kem_keypair(unsigned char *pk, unsigned char *sk) {
+    size_t i;
+    PQCLEAN_KYBER512_CLEAN_indcpa_keypair(pk, sk);
+    for (i = 0; i < KYBER_INDCPA_PUBLICKEYBYTES; i++) {
+        sk[i + KYBER_INDCPA_SECRETKEYBYTES] = pk[i];
+    }
+    hash_h(sk + KYBER_SECRETKEYBYTES - 2 * KYBER_SYMBYTES, pk, KYBER_PUBLICKEYBYTES);
+    randombytes(sk + KYBER_SECRETKEYBYTES - KYBER_SYMBYTES, KYBER_SYMBYTES);    /* Value z for pseudo-random output on reject */
+    return 0;
+}
+
+/*************************************************
+* Name:        crypto_kem_enc
+*
+* Description: Generates cipher text and shared
+*              secret for given public key
+*
+* Arguments:   - unsigned char *ct:       pointer to output cipher text (an already allocated array of CRYPTO_CIPHERTEXTBYTES bytes)
+*              - unsigned char *ss:       pointer to output shared secret (an already allocated array of CRYPTO_BYTES bytes)
+*              - const unsigned char *pk: pointer to input public key (an already allocated array of CRYPTO_PUBLICKEYBYTES bytes)
+*
+* Returns 0 (success)
+**************************************************/
+int PQCLEAN_KYBER512_CLEAN_crypto_kem_enc(unsigned char *ct, unsigned char *ss, const unsigned char *pk) {
+    unsigned char  kr[2 * KYBER_SYMBYTES];                                   /* Will contain key, coins */
+    unsigned char buf[2 * KYBER_SYMBYTES];
+
+    randombytes(buf, KYBER_SYMBYTES);
+    hash_h(buf, buf, KYBER_SYMBYTES);                                        /* Don't release system RNG output */
+
+    hash_h(buf + KYBER_SYMBYTES, pk, KYBER_PUBLICKEYBYTES);                  /* Multitarget countermeasure for coins + contributory KEM */
+    hash_g(kr, buf, 2 * KYBER_SYMBYTES);
+
+    PQCLEAN_KYBER512_CLEAN_indcpa_enc(ct, buf, pk, kr + KYBER_SYMBYTES);                            /* coins are in kr+KYBER_SYMBYTES */
+
+    hash_h(kr + KYBER_SYMBYTES, ct, KYBER_CIPHERTEXTBYTES);                  /* overwrite coins in kr with H(c) */
+    kdf(ss, kr, 2 * KYBER_SYMBYTES);                                         /* hash concatenation of pre-k and H(c) to k */
+    return 0;
+}
+
+/*************************************************
+* Name:        crypto_kem_dec
+*
+* Description: Generates shared secret for given
+*              cipher text and private key
+*
+* Arguments:   - unsigned char *ss:       pointer to output shared secret (an already allocated array of CRYPTO_BYTES bytes)
+*              - const unsigned char *ct: pointer to input cipher text (an already allocated array of CRYPTO_CIPHERTEXTBYTES bytes)
+*              - const unsigned char *sk: pointer to input private key (an already allocated array of CRYPTO_SECRETKEYBYTES bytes)
+*
+* Returns 0.
+*
+* On failure, ss will contain a pseudo-random value.
+**************************************************/
+int PQCLEAN_KYBER512_CLEAN_crypto_kem_dec(unsigned char *ss, const unsigned char *ct, const unsigned char *sk) {
+    size_t i;
+    unsigned char fail;
+    unsigned char cmp[KYBER_CIPHERTEXTBYTES];
+    unsigned char buf[2 * KYBER_SYMBYTES];
+    unsigned char kr[2 * KYBER_SYMBYTES];                                    /* Will contain key, coins */
+    const unsigned char *pk = sk + KYBER_INDCPA_SECRETKEYBYTES;
+
+    PQCLEAN_KYBER512_CLEAN_indcpa_dec(buf, ct, sk);
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {                                   /* Multitarget countermeasure for coins + contributory KEM */
+        buf[KYBER_SYMBYTES + i] = sk[KYBER_SECRETKEYBYTES - 2 * KYBER_SYMBYTES + i];    /* Save hash by storing H(pk) in sk */
+    }
+    hash_g(kr, buf, 2 * KYBER_SYMBYTES);
+
+    PQCLEAN_KYBER512_CLEAN_indcpa_enc(cmp, buf, pk, kr + KYBER_SYMBYTES);                           /* coins are in kr+KYBER_SYMBYTES */
+
+    fail = PQCLEAN_KYBER512_CLEAN_verify(ct, cmp, KYBER_CIPHERTEXTBYTES);
+
+    hash_h(kr + KYBER_SYMBYTES, ct, KYBER_CIPHERTEXTBYTES);                  /* overwrite coins in kr with H(c)  */
+
+    PQCLEAN_KYBER512_CLEAN_cmov(kr, sk + KYBER_SECRETKEYBYTES - KYBER_SYMBYTES, KYBER_SYMBYTES, fail); /* Overwrite pre-k with z on re-encryption failure */
+
+    kdf(ss, kr, 2 * KYBER_SYMBYTES);                                         /* hash concatenation of pre-k and H(c) to k */
+    return 0;
+}
diff --git a/crypto_kem/kyber512/clean/ntt.c b/crypto_kem/kyber512/clean/ntt.c
new file mode 100644
index 00000000..3295aaac
--- /dev/null
+++ b/crypto_kem/kyber512/clean/ntt.c
@@ -0,0 +1,157 @@
+#include "ntt.h"
+
+#include "params.h"
+#include "reduce.h"
+
+#include <stdint.h>
+
+/* Code to generate zetas and zetas_inv used in the number-theoretic transform:
+
+#define KYBER_ROOT_OF_UNITY 17
+
+static const uint16_t tree[128] = {
+  0, 64, 32, 96, 16, 80, 48, 112, 8, 72, 40, 104, 24, 88, 56, 120,
+  4, 68, 36, 100, 20, 84, 52, 116, 12, 76, 44, 108, 28, 92, 60, 124,
+  2, 66, 34, 98, 18, 82, 50, 114, 10, 74, 42, 106, 26, 90, 58, 122,
+  6, 70, 38, 102, 22, 86, 54, 118, 14, 78, 46, 110, 30, 94, 62, 126,
+  1, 65, 33, 97, 17, 81, 49, 113, 9, 73, 41, 105, 25, 89, 57, 121,
+  5, 69, 37, 101, 21, 85, 53, 117, 13, 77, 45, 109, 29, 93, 61, 125,
+  3, 67, 35, 99, 19, 83, 51, 115, 11, 75, 43, 107, 27, 91, 59, 123,
+  7, 71, 39, 103, 23, 87, 55, 119, 15, 79, 47, 111, 31, 95, 63, 127};
+
+
+static int16_t fqmul(int16_t a, int16_t b) {
+  return montgomery_reduce((int32_t)a*b);
+}
+
+void init_ntt() {
+  unsigned int i, j, k;
+  int16_t tmp[128];
+
+  tmp[0] = MONT;
+  for(i = 1; i < 128; ++i)
+    tmp[i] = fqmul(tmp[i-1], KYBER_ROOT_OF_UNITY*MONT % KYBER_Q);
+
+  for(i = 0; i < 128; ++i)
+    zetas[i] = tmp[tree[i]];
+
+  k = 0;
+  for(i = 64; i >= 1; i >>= 1)
+    for(j = i; j < 2*i; ++j)
+      zetas_inv[k++] = -tmp[128 - tree[j]];
+
+  zetas_inv[127] = MONT * (MONT * (KYBER_Q - 1) * ((KYBER_Q - 1)/128) % KYBER_Q) % KYBER_Q;
+}
+
+*/
+int16_t PQCLEAN_KYBER512_CLEAN_zetas[128] = {
+    2285, 2571, 2970, 1812, 1493, 1422, 287, 202, 3158, 622, 1577, 182, 962, 2127, 1855, 1468,
+    573, 2004, 264, 383, 2500, 1458, 1727, 3199, 2648, 1017, 732, 608, 1787, 411, 3124, 1758,
+    1223, 652, 2777, 1015, 2036, 1491, 3047, 1785, 516, 3321, 3009, 2663, 1711, 2167, 126, 1469,
+    2476, 3239, 3058, 830, 107, 1908, 3082, 2378, 2931, 961, 1821, 2604, 448, 2264, 677, 2054,
+    2226, 430, 555, 843, 2078, 871, 1550, 105, 422, 587, 177, 3094, 3038, 2869, 1574, 1653,
+    3083, 778, 1159, 3182, 2552, 1483, 2727, 1119, 1739, 644, 2457, 349, 418, 329, 3173, 3254,
+    817, 1097, 603, 610, 1322, 2044, 1864, 384, 2114, 3193, 1218, 1994, 2455, 220, 2142, 1670,
+    2144, 1799, 2051, 794, 1819, 2475, 2459, 478, 3221, 3021, 996, 991, 958, 1869, 1522, 1628
+};
+
+int16_t PQCLEAN_KYBER512_CLEAN_zetas_inv[128] = {
+    1701, 1807, 1460, 2371, 2338, 2333, 308, 108, 2851, 870, 854, 1510, 2535, 1278, 1530, 1185,
+    1659, 1187, 3109, 874, 1335, 2111, 136, 1215, 2945, 1465, 1285, 2007, 2719, 2726, 2232, 2512,
+    75, 156, 3000, 2911, 2980, 872, 2685, 1590, 2210, 602, 1846, 777, 147, 2170, 2551, 246,
+    1676, 1755, 460, 291, 235, 3152, 2742, 2907, 3224, 1779, 2458, 1251, 2486, 2774, 2899, 1103,
+    1275, 2652, 1065, 2881, 725, 1508, 2368, 398, 951, 247, 1421, 3222, 2499, 271, 90, 853,
+    1860, 3203, 1162, 1618, 666, 320, 8, 2813, 1544, 282, 1838, 1293, 2314, 552, 2677, 2106,
+    1571, 205, 2918, 1542, 2721, 2597, 2312, 681, 130, 1602, 1871, 829, 2946, 3065, 1325, 2756,
+    1861, 1474, 1202, 2367, 3147, 1752, 2707, 171, 3127, 3042, 1907, 1836, 1517, 359, 758, 1441
+};
+
+
+/*************************************************
+* Name:        fqmul
+*
+* Description: Multiplication followed by Montgomery reduction
+*
+* Arguments:   - int16_t a: first factor
+*              - int16_t b: second factor
+*
+* Returns 16-bit integer congruent to a*b*R^{-1} mod q
+**************************************************/
+static int16_t fqmul(int16_t a, int16_t b) {
+    return PQCLEAN_KYBER512_CLEAN_montgomery_reduce((int32_t)a * b);
+}
+
+/*************************************************
+* Name:        ntt
+*
+* Description: Inplace number-theoretic transform (NTT) in Rq
+*              input is in standard order, output is in bitreversed order
+*
+* Arguments:   - int16_t *poly: pointer to input/output vector of 256 elements of Zq
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_ntt(int16_t *poly) {
+    unsigned int len, start, j, k;
+    int16_t t, zeta;
+
+    k = 1;
+    for (len = 128; len >= 2; len >>= 1) {
+        for (start = 0; start < 256; start = j + len) {
+            zeta = PQCLEAN_KYBER512_CLEAN_zetas[k++];
+            for (j = start; j < start + len; ++j) {
+                t = fqmul(zeta, poly[j + len]);
+                poly[j + len] = poly[j] - t;
+                poly[j] = poly[j] + t;
+            }
+        }
+    }
+}
+
+/*************************************************
+* Name:        invntt
+*
+* Description: Inplace inverse number-theoretic transform in Rq
+*              input is in bitreversed order, output is in standard order
+*
+* Arguments:   - int16_t *poly: pointer to input/output vector of 256 elements of Zq
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_invntt(int16_t *poly) {
+    unsigned int start, len, j, k;
+    int16_t t, zeta;
+
+    k = 0;
+    for (len = 2; len <= 128; len <<= 1) {
+        for (start = 0; start < 256; start = j + len) {
+            zeta = PQCLEAN_KYBER512_CLEAN_zetas_inv[k++];
+            for (j = start; j < start + len; ++j) {
+                t = poly[j];
+                poly[j] = PQCLEAN_KYBER512_CLEAN_barrett_reduce(t + poly[j + len]);
+                poly[j + len] = t - poly[j + len];
+                poly[j + len] = fqmul(zeta, poly[j + len]);
+            }
+        }
+    }
+
+    for (j = 0; j < 256; ++j) {
+        poly[j] = fqmul(poly[j], PQCLEAN_KYBER512_CLEAN_zetas_inv[127]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_basemul
+*
+* Description: Multiplication of polynomials in Zq[X]/((X^2-zeta))
+*              used for multiplication of elements in Rq in NTT domain
+*
+* Arguments:   - int16_t r[2]: pointer to the output polynomial
+*              - const int16_t a[2]: pointer to the first factor
+*              - const int16_t b[2]: pointer to the second factor
+*              - int16_t zeta: integer defining the reduction polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_basemul(int16_t r[2], const int16_t a[2], const int16_t b[2], int16_t zeta) {
+    r[0]  = fqmul(a[1], b[1]);
+    r[0]  = fqmul(r[0], zeta);
+    r[0] += fqmul(a[0], b[0]);
+
+    r[1]  = fqmul(a[0], b[1]);
+    r[1] += fqmul(a[1], b[0]);
+}
diff --git a/crypto_kem/kyber512/clean/ntt.h b/crypto_kem/kyber512/clean/ntt.h
new file mode 100644
index 00000000..678c311d
--- /dev/null
+++ b/crypto_kem/kyber512/clean/ntt.h
@@ -0,0 +1,13 @@
+#ifndef NTT_H
+#define NTT_H
+
+#include <stdint.h>
+
+extern int16_t PQCLEAN_KYBER512_CLEAN_zetas[128];
+extern int16_t PQCLEAN_KYBER512_CLEAN_zetas_inv[128];
+
+void PQCLEAN_KYBER512_CLEAN_ntt(int16_t *poly);
+void PQCLEAN_KYBER512_CLEAN_invntt(int16_t *poly);
+void PQCLEAN_KYBER512_CLEAN_basemul(int16_t r[2], const int16_t a[2], const int16_t b[2], int16_t zeta);
+
+#endif
diff --git a/crypto_kem/kyber512/clean/params.h b/crypto_kem/kyber512/clean/params.h
new file mode 100644
index 00000000..b9865f7b
--- /dev/null
+++ b/crypto_kem/kyber512/clean/params.h
@@ -0,0 +1,31 @@
+#ifndef PARAMS_H
+#define PARAMS_H
+
+#define KYBER_K 2 /* Change this for different security strengths */
+
+/* Don't change parameters below this line */
+
+#define KYBER_N 256
+#define KYBER_Q 3329
+
+#define KYBER_ETA 2
+
+#define KYBER_SYMBYTES 32   /* size in bytes of hashes, and seeds */
+#define KYBER_SSBYTES  32   /* size in bytes of shared key */
+
+#define KYBER_POLYBYTES              384
+#define KYBER_POLYVECBYTES           (KYBER_K * KYBER_POLYBYTES)
+
+#define KYBER_POLYCOMPRESSEDBYTES    96
+#define KYBER_POLYVECCOMPRESSEDBYTES (KYBER_K * 320)
+
+#define KYBER_INDCPA_MSGBYTES       KYBER_SYMBYTES
+#define KYBER_INDCPA_PUBLICKEYBYTES (KYBER_POLYVECBYTES + KYBER_SYMBYTES)
+#define KYBER_INDCPA_SECRETKEYBYTES (KYBER_POLYVECBYTES)
+#define KYBER_INDCPA_BYTES          (KYBER_POLYVECCOMPRESSEDBYTES + KYBER_POLYCOMPRESSEDBYTES)
+
+#define KYBER_PUBLICKEYBYTES  (KYBER_INDCPA_PUBLICKEYBYTES)
+#define KYBER_SECRETKEYBYTES  (KYBER_INDCPA_SECRETKEYBYTES +  KYBER_INDCPA_PUBLICKEYBYTES + 2*KYBER_SYMBYTES) /* 32 bytes of additional space to save H(pk) */
+#define KYBER_CIPHERTEXTBYTES  KYBER_INDCPA_BYTES
+
+#endif
diff --git a/crypto_kem/kyber512/clean/poly.c b/crypto_kem/kyber512/clean/poly.c
new file mode 100644
index 00000000..f72508d9
--- /dev/null
+++ b/crypto_kem/kyber512/clean/poly.c
@@ -0,0 +1,287 @@
+#include "poly.h"
+
+#include "cbd.h"
+#include "ntt.h"
+#include "params.h"
+#include "reduce.h"
+#include "symmetric.h"
+
+#include <stdint.h>
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_compress
+*
+* Description: Compression and subsequent serialization of a polynomial
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYCOMPRESSEDBYTES bytes)
+*              - const poly *a:    pointer to input polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_compress(unsigned char *r, poly *a) {
+    uint8_t t[8];
+    int i, j, k = 0;
+
+    PQCLEAN_KYBER512_CLEAN_poly_csubq(a);
+
+    for (i = 0; i < KYBER_N; i += 8) {
+        for (j = 0; j < 8; j++) {
+            t[j] = ((((uint32_t)a->coeffs[i + j] << 3) + KYBER_Q / 2) / KYBER_Q) & 7;
+        }
+
+        r[k]   =  t[0]       | (t[1] << 3) | (t[2] << 6);
+        r[k + 1] = (t[2] >> 2) | (t[3] << 1) | (t[4] << 4) | (t[5] << 7);
+        r[k + 2] = (t[5] >> 1) | (t[6] << 2) | (t[7] << 5);
+        k += 3;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_decompress
+*
+* Description: De-serialization and subsequent decompression of a polynomial;
+*              approximate inverse of PQCLEAN_KYBER512_CLEAN_poly_compress
+*
+* Arguments:   - poly *r:                pointer to output polynomial
+*              - const unsigned char *a: pointer to input byte array (of length KYBER_POLYCOMPRESSEDBYTES bytes)
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_decompress(poly *r, const unsigned char *a) {
+    int i;
+    for (i = 0; i < KYBER_N; i += 8) {
+        r->coeffs[i + 0] =  (((a[0] & 7) * KYBER_Q) + 4) >> 3;
+        r->coeffs[i + 1] = ((((a[0] >> 3) & 7) * KYBER_Q) + 4) >> 3;
+        r->coeffs[i + 2] = ((((a[0] >> 6) | ((a[1] << 2) & 4)) * KYBER_Q) + 4) >> 3;
+        r->coeffs[i + 3] = ((((a[1] >> 1) & 7) * KYBER_Q) + 4) >> 3;
+        r->coeffs[i + 4] = ((((a[1] >> 4) & 7) * KYBER_Q) + 4) >> 3;
+        r->coeffs[i + 5] = ((((a[1] >> 7) | ((a[2] << 1) & 6)) * KYBER_Q) + 4) >> 3;
+        r->coeffs[i + 6] = ((((a[2] >> 2) & 7) * KYBER_Q) + 4) >> 3;
+        r->coeffs[i + 7] = ((((a[2] >> 5)) * KYBER_Q) + 4) >> 3;
+        a += 3;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_tobytes
+*
+* Description: Serialization of a polynomial
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYBYTES bytes)
+*              - const poly *a:    pointer to input polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_tobytes(unsigned char *r, poly *a) {
+    int i;
+    uint16_t t0, t1;
+
+    PQCLEAN_KYBER512_CLEAN_poly_csubq(a);
+
+    for (i = 0; i < KYBER_N / 2; i++) {
+        t0 = a->coeffs[2 * i];
+        t1 = a->coeffs[2 * i + 1];
+        r[3 * i] = t0 & 0xff;
+        r[3 * i + 1] = (t0 >> 8) | ((t1 & 0xf) << 4);
+        r[3 * i + 2] = (t1 >> 4) & 0xff;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_frombytes
+*
+* Description: De-serialization of a polynomial;
+*              inverse of PQCLEAN_KYBER512_CLEAN_poly_tobytes
+*
+* Arguments:   - poly *r:                pointer to output polynomial
+*              - const unsigned char *a: pointer to input byte array (of KYBER_POLYBYTES bytes)
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_frombytes(poly *r, const unsigned char *a) {
+    int i;
+
+    for (i = 0; i < KYBER_N / 2; i++) {
+        r->coeffs[2 * i]   = a[3 * i]        | ((uint16_t)a[3 * i + 1] & 0x0f) << 8;
+        r->coeffs[2 * i + 1] = a[3 * i + 1] >> 4 | ((uint16_t)a[3 * i + 2] & 0xff) << 4;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_getnoise
+*
+* Description: Sample a polynomial deterministically from a seed and a nonce,
+*              with output polynomial close to centered binomial distribution
+*              with parameter KYBER_ETA
+*
+* Arguments:   - poly *r:                   pointer to output polynomial
+*              - const unsigned char *seed: pointer to input seed (pointing to array of length KYBER_SYMBYTES bytes)
+*              - unsigned char nonce:       one-byte input nonce
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_getnoise(poly *r, const unsigned char *seed, unsigned char nonce) {
+    unsigned char buf[KYBER_ETA * KYBER_N / 4];
+
+    prf(buf, KYBER_ETA * KYBER_N / 4, seed, nonce);
+    PQCLEAN_KYBER512_CLEAN_cbd(r, buf);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_ntt
+*
+* Description: Computes negacyclic number-theoretic transform (NTT) of
+*              a polynomial in place;
+*              inputs assumed to be in normal order, output in bitreversed order
+*
+* Arguments:   - uint16_t *r: pointer to in/output polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_ntt(poly *r) {
+    PQCLEAN_KYBER512_CLEAN_ntt(r->coeffs);
+    PQCLEAN_KYBER512_CLEAN_poly_reduce(r);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_invntt
+*
+* Description: Computes inverse of negacyclic number-theoretic transform (NTT) of
+*              a polynomial in place;
+*              inputs assumed to be in bitreversed order, output in normal order
+*
+* Arguments:   - uint16_t *a: pointer to in/output polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_invntt(poly *r) {
+    PQCLEAN_KYBER512_CLEAN_invntt(r->coeffs);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_basemul
+*
+* Description: Multiplication of two polynomials in NTT domain
+*
+* Arguments:   - poly *r:       pointer to output polynomial
+*              - const poly *a: pointer to first input polynomial
+*              - const poly *b: pointer to second input polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_basemul(poly *r, const poly *a, const poly *b) {
+    unsigned int i;
+
+    for (i = 0; i < KYBER_N / 4; ++i) {
+        PQCLEAN_KYBER512_CLEAN_basemul(r->coeffs + 4 * i, a->coeffs + 4 * i, b->coeffs + 4 * i, PQCLEAN_KYBER512_CLEAN_zetas[64 + i]);
+        PQCLEAN_KYBER512_CLEAN_basemul(r->coeffs + 4 * i + 2, a->coeffs + 4 * i + 2, b->coeffs + 4 * i + 2, - PQCLEAN_KYBER512_CLEAN_zetas[64 + i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_frommont
+*
+* Description: Inplace conversion of all coefficients of a polynomial
+*              from Montgomery domain to normal domain
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_frommont(poly *r) {
+    int i;
+    const int16_t f = (1ULL << 32) % KYBER_Q;
+
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = PQCLEAN_KYBER512_CLEAN_montgomery_reduce((int32_t)r->coeffs[i] * f);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_reduce
+*
+* Description: Applies Barrett reduction to all coefficients of a polynomial
+*              for details of the Barrett reduction see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_reduce(poly *r) {
+    int i;
+
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = PQCLEAN_KYBER512_CLEAN_barrett_reduce(r->coeffs[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_csubq
+*
+* Description: Applies conditional subtraction of q to each coefficient of a polynomial
+*              for details of conditional subtraction of q see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_csubq(poly *r) {
+    int i;
+
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = PQCLEAN_KYBER512_CLEAN_csubq(r->coeffs[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_add
+*
+* Description: Add two polynomials
+*
+* Arguments: - poly *r:       pointer to output polynomial
+*            - const poly *a: pointer to first input polynomial
+*            - const poly *b: pointer to second input polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_add(poly *r, const poly *a, const poly *b) {
+    int i;
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = a->coeffs[i] + b->coeffs[i];
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_sub
+*
+* Description: Subtract two polynomials
+*
+* Arguments: - poly *r:       pointer to output polynomial
+*            - const poly *a: pointer to first input polynomial
+*            - const poly *b: pointer to second input polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_sub(poly *r, const poly *a, const poly *b) {
+    int i;
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = a->coeffs[i] - b->coeffs[i];
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_frommsg
+*
+* Description: Convert 32-byte message to polynomial
+*
+* Arguments:   - poly *r:                  pointer to output polynomial
+*              - const unsigned char *msg: pointer to input message
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_frommsg(poly *r, const unsigned char msg[KYBER_SYMBYTES]) {
+    int i, j;
+    uint16_t mask;
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        for (j = 0; j < 8; j++) {
+            mask = -((msg[i] >> j) & 1);
+            r->coeffs[8 * i + j] = mask & ((KYBER_Q + 1) / 2);
+        }
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_poly_tomsg
+*
+* Description: Convert polynomial to 32-byte message
+*
+* Arguments:   - unsigned char *msg: pointer to output message
+*              - const poly *a:      pointer to input polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_poly_tomsg(unsigned char msg[KYBER_SYMBYTES], poly *a) {
+    uint16_t t;
+    int i, j;
+
+    PQCLEAN_KYBER512_CLEAN_poly_csubq(a);
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        msg[i] = 0;
+        for (j = 0; j < 8; j++) {
+            t = (((a->coeffs[8 * i + j] << 1) + KYBER_Q / 2) / KYBER_Q) & 1;
+            msg[i] |= t << j;
+        }
+    }
+}
diff --git a/crypto_kem/kyber512/clean/poly.h b/crypto_kem/kyber512/clean/poly.h
new file mode 100644
index 00000000..9002a530
--- /dev/null
+++ b/crypto_kem/kyber512/clean/poly.h
@@ -0,0 +1,38 @@
+#ifndef POLY_H
+#define POLY_H
+
+#include "params.h"
+
+#include <stdint.h>
+
+/*
+ * Elements of R_q = Z_q[X]/(X^n + 1). Represents polynomial
+ * coeffs[0] + X*coeffs[1] + X^2*xoeffs[2] + ... + X^{n-1}*coeffs[n-1]
+ */
+typedef struct {
+    int16_t coeffs[KYBER_N];
+} poly;
+
+void PQCLEAN_KYBER512_CLEAN_poly_compress(unsigned char *r, poly *a);
+void PQCLEAN_KYBER512_CLEAN_poly_decompress(poly *r, const unsigned char *a);
+
+void PQCLEAN_KYBER512_CLEAN_poly_tobytes(unsigned char *r, poly *a);
+void PQCLEAN_KYBER512_CLEAN_poly_frombytes(poly *r, const unsigned char *a);
+
+void PQCLEAN_KYBER512_CLEAN_poly_frommsg(poly *r, const unsigned char msg[KYBER_SYMBYTES]);
+void PQCLEAN_KYBER512_CLEAN_poly_tomsg(unsigned char msg[KYBER_SYMBYTES], poly *a);
+
+void PQCLEAN_KYBER512_CLEAN_poly_getnoise(poly *r, const unsigned char *seed, unsigned char nonce);
+
+void PQCLEAN_KYBER512_CLEAN_poly_ntt(poly *r);
+void PQCLEAN_KYBER512_CLEAN_poly_invntt(poly *r);
+void PQCLEAN_KYBER512_CLEAN_poly_basemul(poly *r, const poly *a, const poly *b);
+void PQCLEAN_KYBER512_CLEAN_poly_frommont(poly *r);
+
+void PQCLEAN_KYBER512_CLEAN_poly_reduce(poly *r);
+void PQCLEAN_KYBER512_CLEAN_poly_csubq(poly *r);
+
+void PQCLEAN_KYBER512_CLEAN_poly_add(poly *r, const poly *a, const poly *b);
+void PQCLEAN_KYBER512_CLEAN_poly_sub(poly *r, const poly *a, const poly *b);
+
+#endif
diff --git a/crypto_kem/kyber512/clean/polyvec.c b/crypto_kem/kyber512/clean/polyvec.c
new file mode 100644
index 00000000..2dbed8aa
--- /dev/null
+++ b/crypto_kem/kyber512/clean/polyvec.c
@@ -0,0 +1,186 @@
+#include "polyvec.h"
+
+#include "poly.h"
+
+#include <stdint.h>
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_polyvec_compress
+*
+* Description: Compress and serialize vector of polynomials
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYVECCOMPRESSEDBYTES)
+*              - const polyvec *a: pointer to input vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_polyvec_compress(unsigned char *r, polyvec *a) {
+    int i, j, k;
+
+    PQCLEAN_KYBER512_CLEAN_polyvec_csubq(a);
+
+    uint16_t t[4];
+    for (i = 0; i < KYBER_K; i++) {
+        for (j = 0; j < KYBER_N / 4; j++) {
+            for (k = 0; k < 4; k++) {
+                t[k] = ((((uint32_t)a->vec[i].coeffs[4 * j + k] << 10) + KYBER_Q / 2) / KYBER_Q) & 0x3ff;
+            }
+
+            r[5 * j + 0] =  t[0] & 0xff;
+            r[5 * j + 1] = (t[0] >>  8) | ((t[1] & 0x3f) << 2);
+            r[5 * j + 2] = ((t[1] >>  6) | ((t[2] & 0x0f) << 4)) & 0xff;
+            r[5 * j + 3] = ((t[2] >>  4) | ((t[3] & 0x03) << 6)) & 0xff;
+            r[5 * j + 4] = (t[3] >>  2) & 0xff;
+        }
+        r += 320;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_polyvec_decompress
+*
+* Description: De-serialize and decompress vector of polynomials;
+*              approximate inverse of PQCLEAN_KYBER512_CLEAN_polyvec_compress
+*
+* Arguments:   - polyvec *r:       pointer to output vector of polynomials
+*              - unsigned char *a: pointer to input byte array (of length KYBER_POLYVECCOMPRESSEDBYTES)
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_polyvec_decompress(polyvec *r, const unsigned char *a) {
+    int i, j;
+    for (i = 0; i < KYBER_K; i++) {
+        for (j = 0; j < KYBER_N / 4; j++) {
+            r->vec[i].coeffs[4 * j + 0] =  (((a[5 * j + 0]       | (((uint32_t)a[5 * j + 1] & 0x03) << 8)) * KYBER_Q) + 512) >> 10;
+            r->vec[i].coeffs[4 * j + 1] = ((((a[5 * j + 1] >> 2) | (((uint32_t)a[5 * j + 2] & 0x0f) << 6)) * KYBER_Q) + 512) >> 10;
+            r->vec[i].coeffs[4 * j + 2] = ((((a[5 * j + 2] >> 4) | (((uint32_t)a[5 * j + 3] & 0x3f) << 4)) * KYBER_Q) + 512) >> 10;
+            r->vec[i].coeffs[4 * j + 3] = ((((a[5 * j + 3] >> 6) | (((uint32_t)a[5 * j + 4] & 0xff) << 2)) * KYBER_Q) + 512) >> 10;
+        }
+        a += 320;
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_polyvec_tobytes
+*
+* Description: Serialize vector of polynomials
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYVECBYTES)
+*              - const polyvec *a: pointer to input vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_polyvec_tobytes(unsigned char *r, polyvec *a) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_tobytes(r + i * KYBER_POLYBYTES, &a->vec[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_polyvec_frombytes
+*
+* Description: De-serialize vector of polynomials;
+*              inverse of PQCLEAN_KYBER512_CLEAN_polyvec_tobytes
+*
+* Arguments:   - unsigned char *r: pointer to output byte array
+*              - const polyvec *a: pointer to input vector of polynomials (of length KYBER_POLYVECBYTES)
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_polyvec_frombytes(polyvec *r, const unsigned char *a) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_frombytes(&r->vec[i], a + i * KYBER_POLYBYTES);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_polyvec_ntt
+*
+* Description: Apply forward NTT to all elements of a vector of polynomials
+*
+* Arguments:   - polyvec *r: pointer to in/output vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_polyvec_ntt(polyvec *r) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_ntt(&r->vec[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_polyvec_invntt
+*
+* Description: Apply inverse NTT to all elements of a vector of polynomials
+*
+* Arguments:   - polyvec *r: pointer to in/output vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_polyvec_invntt(polyvec *r) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_invntt(&r->vec[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_polyvec_pointwise_acc
+*
+* Description: Pointwise multiply elements of a and b and accumulate into r
+*
+* Arguments: - poly *r:          pointer to output polynomial
+*            - const polyvec *a: pointer to first input vector of polynomials
+*            - const polyvec *b: pointer to second input vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_polyvec_pointwise_acc(poly *r, const polyvec *a, const polyvec *b) {
+    int i;
+    poly t;
+
+    PQCLEAN_KYBER512_CLEAN_poly_basemul(r, &a->vec[0], &b->vec[0]);
+    for (i = 1; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_basemul(&t, &a->vec[i], &b->vec[i]);
+        PQCLEAN_KYBER512_CLEAN_poly_add(r, r, &t);
+    }
+
+    PQCLEAN_KYBER512_CLEAN_poly_reduce(r);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_polyvec_reduce
+*
+* Description: Applies Barrett reduction to each coefficient
+*              of each element of a vector of polynomials
+*              for details of the Barrett reduction see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_polyvec_reduce(polyvec *r) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_reduce(&r->vec[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_polyvec_csubq
+*
+* Description: Applies conditional subtraction of q to each coefficient
+*              of each element of a vector of polynomials
+*              for details of conditional subtraction of q see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_polyvec_csubq(polyvec *r) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_csubq(&r->vec[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_polyvec_add
+*
+* Description: Add vectors of polynomials
+*
+* Arguments: - polyvec *r:       pointer to output vector of polynomials
+*            - const polyvec *a: pointer to first input vector of polynomials
+*            - const polyvec *b: pointer to second input vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_polyvec_add(polyvec *r, const polyvec *a, const polyvec *b) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER512_CLEAN_poly_add(&r->vec[i], &a->vec[i], &b->vec[i]);
+    }
+}
diff --git a/crypto_kem/kyber512/clean/polyvec.h b/crypto_kem/kyber512/clean/polyvec.h
new file mode 100644
index 00000000..5e2ed31f
--- /dev/null
+++ b/crypto_kem/kyber512/clean/polyvec.h
@@ -0,0 +1,27 @@
+#ifndef POLYVEC_H
+#define POLYVEC_H
+
+#include "params.h"
+#include "poly.h"
+
+typedef struct {
+    poly vec[KYBER_K];
+} polyvec;
+
+void PQCLEAN_KYBER512_CLEAN_polyvec_compress(unsigned char *r, polyvec *a);
+void PQCLEAN_KYBER512_CLEAN_polyvec_decompress(polyvec *r, const unsigned char *a);
+
+void PQCLEAN_KYBER512_CLEAN_polyvec_tobytes(unsigned char *r, polyvec *a);
+void PQCLEAN_KYBER512_CLEAN_polyvec_frombytes(polyvec *r, const unsigned char *a);
+
+void PQCLEAN_KYBER512_CLEAN_polyvec_ntt(polyvec *r);
+void PQCLEAN_KYBER512_CLEAN_polyvec_invntt(polyvec *r);
+
+void PQCLEAN_KYBER512_CLEAN_polyvec_pointwise_acc(poly *r, const polyvec *a, const polyvec *b);
+
+void PQCLEAN_KYBER512_CLEAN_polyvec_reduce(polyvec *r);
+void PQCLEAN_KYBER512_CLEAN_polyvec_csubq(polyvec *r);
+
+void PQCLEAN_KYBER512_CLEAN_polyvec_add(polyvec *r, const polyvec *a, const polyvec *b);
+
+#endif
diff --git a/crypto_kem/kyber512/clean/reduce.c b/crypto_kem/kyber512/clean/reduce.c
new file mode 100644
index 00000000..ecc99dd4
--- /dev/null
+++ b/crypto_kem/kyber512/clean/reduce.c
@@ -0,0 +1,62 @@
+#include "reduce.h"
+
+#include "params.h"
+
+#include <stdint.h>
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_montgomery_reduce
+*
+* Description: Montgomery reduction; given a 32-bit integer a, computes
+*              16-bit integer congruent to a * R^-1 mod q,
+*              where R=2^16
+*
+* Arguments:   - int32_t a: input integer to be reduced; has to be in {-q2^15,...,q2^15-1}
+*
+* Returns:     integer in {-q+1,...,q-1} congruent to a * R^-1 modulo q.
+**************************************************/
+int16_t PQCLEAN_KYBER512_CLEAN_montgomery_reduce(int32_t a) {
+    int32_t t;
+    int16_t u;
+
+    u = (int16_t)(a * QINV);
+    t = (int32_t)u * KYBER_Q;
+    t = a - t;
+    t >>= 16;
+    return (int16_t)t;
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_barrett_reduce
+*
+* Description: Barrett reduction; given a 16-bit integer a, computes
+*              16-bit integer congruent to a mod q in {0,...,q}
+*
+* Arguments:   - int16_t a: input integer to be reduced
+*
+* Returns:     integer in {0,...,q} congruent to a modulo q.
+**************************************************/
+int16_t PQCLEAN_KYBER512_CLEAN_barrett_reduce(int16_t a) {
+    int32_t t;
+    const int32_t v = (1U << 26) / KYBER_Q + 1;
+
+    t = v * a;
+    t >>= 26;
+    t *= KYBER_Q;
+    return (int16_t)(a - t);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_csubq
+*
+* Description: Conditionallly subtract q
+*
+* Arguments:   - int16_t x: input integer
+*
+* Returns:     a - q if a >= q, else a
+**************************************************/
+int16_t PQCLEAN_KYBER512_CLEAN_csubq(int16_t a) {
+    a -= KYBER_Q;
+    a += (a >> 15) & KYBER_Q;
+    return a;
+}
diff --git a/crypto_kem/kyber512/clean/reduce.h b/crypto_kem/kyber512/clean/reduce.h
new file mode 100644
index 00000000..68a7f570
--- /dev/null
+++ b/crypto_kem/kyber512/clean/reduce.h
@@ -0,0 +1,15 @@
+#ifndef REDUCE_H
+#define REDUCE_H
+
+#include <stdint.h>
+
+#define MONT 2285 // 2^16 % Q
+#define QINV 62209 // q^(-1) mod 2^16
+
+int16_t PQCLEAN_KYBER512_CLEAN_montgomery_reduce(int32_t a);
+
+int16_t PQCLEAN_KYBER512_CLEAN_barrett_reduce(int16_t a);
+
+int16_t PQCLEAN_KYBER512_CLEAN_csubq(int16_t a);
+
+#endif
diff --git a/crypto_kem/kyber512/clean/symmetric-fips202.c b/crypto_kem/kyber512/clean/symmetric-fips202.c
new file mode 100644
index 00000000..345cebb4
--- /dev/null
+++ b/crypto_kem/kyber512/clean/symmetric-fips202.c
@@ -0,0 +1,64 @@
+#include "fips202.h"
+#include "symmetric.h"
+
+#include <stdlib.h>
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_kyber_shake128_absorb
+*
+* Description: Absorb step of the SHAKE128 specialized for the Kyber context.
+*
+* Arguments:   - uint64_t *s:                     pointer to (uninitialized) output Keccak state
+*              - const unsigned char *input:      pointer to KYBER_SYMBYTES input to be absorbed into s
+*              - unsigned char i                  additional byte of input
+*              - unsigned char j                  additional byte of input
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_kyber_shake128_absorb(keccak_state *s, const unsigned char *input, unsigned char x, unsigned char y) {
+    unsigned char extseed[KYBER_SYMBYTES + 2];
+    int i;
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        extseed[i] = input[i];
+    }
+    extseed[i++] = x;
+    extseed[i]   = y;
+    shake128_absorb(s->s, extseed, KYBER_SYMBYTES + 2);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_kyber_shake128_squeezeblocks
+*
+* Description: Squeeze step of SHAKE128 XOF. Squeezes full blocks of SHAKE128_RATE bytes each.
+*              Modifies the state. Can be called multiple times to keep squeezing,
+*              i.e., is incremental.
+*
+* Arguments:   - unsigned char *output:      pointer to output blocks
+*              - size_t nblocks:             number of blocks to be squeezed (written to output)
+*              - keccak_state *s:            pointer to in/output Keccak state
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_kyber_shake128_squeezeblocks(unsigned char *output, size_t nblocks, keccak_state *s) {
+    shake128_squeezeblocks(output, nblocks, s->s);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER512_CLEAN_shake256_prf
+*
+* Description: Usage of SHAKE256 as a PRF, concatenates secret and public input
+*              and then generates outlen bytes of SHAKE256 output
+*
+* Arguments:   - unsigned char *output:      pointer to output
+*              - size_t outlen:              number of requested output bytes
+*              - const unsigned char * key:  pointer to the key (of length KYBER_SYMBYTES)
+*              - const unsigned char nonce:  single-byte nonce (public PRF input)
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_shake256_prf(unsigned char *output, size_t outlen, const unsigned char *key, unsigned char nonce) {
+    unsigned char extkey[KYBER_SYMBYTES + 1];
+    size_t i;
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        extkey[i] = key[i];
+    }
+    extkey[i] = nonce;
+
+    shake256(output, outlen, extkey, KYBER_SYMBYTES + 1);
+}
diff --git a/crypto_kem/kyber512/clean/symmetric.h b/crypto_kem/kyber512/clean/symmetric.h
new file mode 100644
index 00000000..b0f61aeb
--- /dev/null
+++ b/crypto_kem/kyber512/clean/symmetric.h
@@ -0,0 +1,26 @@
+#ifndef SYMMETRIC_H
+#define SYMMETRIC_H
+
+#include "fips202.h"
+#include "params.h"
+
+typedef struct {
+    uint64_t s[25];
+} keccak_state;
+
+void PQCLEAN_KYBER512_CLEAN_kyber_shake128_absorb(keccak_state *s, const unsigned char *input, unsigned char x, unsigned char y);
+void PQCLEAN_KYBER512_CLEAN_kyber_shake128_squeezeblocks(unsigned char *output, size_t nblocks, keccak_state *s);
+void PQCLEAN_KYBER512_CLEAN_shake256_prf(unsigned char *output, size_t outlen, const unsigned char *key, unsigned char nonce);
+
+#define hash_h(OUT, IN, INBYTES) sha3_256(OUT, IN, INBYTES)
+#define hash_g(OUT, IN, INBYTES) sha3_512(OUT, IN, INBYTES)
+#define xof_absorb(STATE, IN, X, Y) PQCLEAN_KYBER512_CLEAN_kyber_shake128_absorb(STATE, IN, X, Y)
+#define xof_squeezeblocks(OUT, OUTBLOCKS, STATE) PQCLEAN_KYBER512_CLEAN_kyber_shake128_squeezeblocks(OUT, OUTBLOCKS, STATE)
+#define prf(OUT, OUTBYTES, KEY, NONCE) PQCLEAN_KYBER512_CLEAN_shake256_prf(OUT, OUTBYTES, KEY, NONCE)
+#define kdf(OUT, IN, INBYTES) shake256(OUT, KYBER_SSBYTES, IN, INBYTES)
+
+#define XOF_BLOCKBYTES 168
+
+typedef keccak_state xof_state;
+
+#endif /* SYMMETRIC_H */
diff --git a/crypto_kem/kyber512/clean/verify.c b/crypto_kem/kyber512/clean/verify.c
new file mode 100644
index 00000000..b8f4dcff
--- /dev/null
+++ b/crypto_kem/kyber512/clean/verify.c
@@ -0,0 +1,51 @@
+#include "verify.h"
+
+#include <stdint.h>
+#include <stdlib.h>
+
+/*************************************************
+* Name:        verify
+*
+* Description: Compare two arrays for equality in constant time.
+*
+* Arguments:   const unsigned char *a: pointer to first byte array
+*              const unsigned char *b: pointer to second byte array
+*              size_t len:             length of the byte arrays
+*
+* Returns 0 if the byte arrays are equal, 1 otherwise
+**************************************************/
+unsigned char PQCLEAN_KYBER512_CLEAN_verify(const unsigned char *a, const unsigned char *b, size_t len) {
+    uint64_t r;
+    size_t i;
+
+    r = 0;
+    for (i = 0; i < len; i++) {
+        r |= a[i] ^ b[i];
+    }
+
+    r = (~r + 1); // Two's complement
+    r >>= 63;
+    return (unsigned char)r;
+}
+
+/*************************************************
+* Name:        cmov
+*
+* Description: Copy len bytes from x to r if b is 1;
+*              don't modify x if b is 0. Requires b to be in {0,1};
+*              assumes two's complement representation of negative integers.
+*              Runs in constant time.
+*
+* Arguments:   unsigned char *r:       pointer to output byte array
+*              const unsigned char *x: pointer to input byte array
+*              size_t len:             Amount of bytes to be copied
+*              unsigned char b:        Condition bit; has to be in {0,1}
+**************************************************/
+void PQCLEAN_KYBER512_CLEAN_cmov(unsigned char *r, const unsigned char *x, size_t len, unsigned char b) {
+    size_t i;
+
+    b = -b;
+    for (i = 0; i < len; i++) {
+        r[i] ^= b & (x[i] ^ r[i]);
+    }
+}
diff --git a/crypto_kem/kyber512/clean/verify.h b/crypto_kem/kyber512/clean/verify.h
new file mode 100644
index 00000000..f2b6bbbc
--- /dev/null
+++ b/crypto_kem/kyber512/clean/verify.h
@@ -0,0 +1,10 @@
+#ifndef VERIFY_H
+#define VERIFY_H
+
+#include <stdio.h>
+
+unsigned char PQCLEAN_KYBER512_CLEAN_verify(const unsigned char *a, const unsigned char *b, size_t len);
+
+void PQCLEAN_KYBER512_CLEAN_cmov(unsigned char *r, const unsigned char *x, size_t len, unsigned char b);
+
+#endif
diff --git a/crypto_kem/kyber768/META.yml b/crypto_kem/kyber768/META.yml
index 8a183ecc..2d8a309e 100644
--- a/crypto_kem/kyber768/META.yml
+++ b/crypto_kem/kyber768/META.yml
@@ -1,9 +1,9 @@
 name: Kyber768
 type: kem
 claimed-nist-level: 3
-length-public-key: 1088
+length-public-key: 1184
+length-ciphertext: 1088
 length-secret-key: 2400
-length-ciphertext: 1152
 length-shared-secret: 32
 nistkat-sha256: d6dbb9399d1ba4ee2d986de3e54a461256b91d6c2f9b90ad2410cf41e09b64d1
 principal-submitter: Peter Schwabe
@@ -19,4 +19,4 @@ auxiliary-submitters:
   - Damien Stehlé
 implementations:
     - name: clean
-      version: https://github.com/pq-crystals/kyber/commit/ab996e7460e5356b0e23aa034e7c2fe6922e60e6
+      version: https://github.com/pq-crystals/kyber/commit/46e283ab575ec92dfe82fb12229ae2d9d6246682
diff --git a/crypto_kem/kyber768/clean/Makefile b/crypto_kem/kyber768/clean/Makefile
index d408dca7..c98f6e49 100644
--- a/crypto_kem/kyber768/clean/Makefile
+++ b/crypto_kem/kyber768/clean/Makefile
@@ -1,8 +1,8 @@
 # This Makefile can be used with GNU Make or BSD Make
 
 LIB=libkyber768_clean.a
-HEADERS=api.h cbd.h indcpa.h ntt.h params.h poly.h polyvec.h reduce.h verify.h
-OBJECTS=cbd.o indcpa.o kem.o ntt.o poly.o polyvec.o precomp.o reduce.o verify.o
+HEADERS=api.h cbd.h indcpa.h ntt.h params.h poly.h polyvec.h reduce.h verify.h symmetric.h
+OBJECTS=cbd.o indcpa.o kem.o ntt.o poly.o polyvec.o reduce.o verify.o symmetric-fips202.o
 
 CFLAGS=-O3 -Wall -Wextra -Wpedantic -Wvla -Werror -Wmissing-prototypes -std=c99 -I../../../common $(EXTRAFLAGS)
 
diff --git a/crypto_kem/kyber768/clean/Makefile.Microsoft_nmake b/crypto_kem/kyber768/clean/Makefile.Microsoft_nmake
index eb7bb622..c037a87e 100644
--- a/crypto_kem/kyber768/clean/Makefile.Microsoft_nmake
+++ b/crypto_kem/kyber768/clean/Makefile.Microsoft_nmake
@@ -2,7 +2,7 @@
 #    nmake /f Makefile.Microsoft_nmake
 
 LIBRARY=libkyber768_clean.lib
-OBJECTS=cbd.obj indcpa.obj kem.obj ntt.obj poly.obj polyvec.obj precomp.obj reduce.obj verify.obj
+OBJECTS=cbd.obj indcpa.obj kem.obj ntt.obj poly.obj polyvec.obj reduce.obj verify.obj symmetric-fips202.obj
 
 CFLAGS=/nologo /I ..\..\..\common /W4 /WX
 
diff --git a/crypto_kem/kyber768/clean/api.h b/crypto_kem/kyber768/clean/api.h
index be974161..35823876 100644
--- a/crypto_kem/kyber768/clean/api.h
+++ b/crypto_kem/kyber768/clean/api.h
@@ -1,19 +1,18 @@
 #ifndef PQCLEAN_KYBER768_CLEAN_API_H
 #define PQCLEAN_KYBER768_CLEAN_API_H
 
-#include <stdint.h>
-
-#define PQCLEAN_KYBER768_CLEAN_CRYPTO_SECRETKEYBYTES 2400
-#define PQCLEAN_KYBER768_CLEAN_CRYPTO_PUBLICKEYBYTES 1088
-#define PQCLEAN_KYBER768_CLEAN_CRYPTO_CIPHERTEXTBYTES 1152
-#define PQCLEAN_KYBER768_CLEAN_CRYPTO_BYTES 32
+#define PQCLEAN_KYBER768_CLEAN_CRYPTO_SECRETKEYBYTES  2400
+#define PQCLEAN_KYBER768_CLEAN_CRYPTO_PUBLICKEYBYTES  1184
+#define PQCLEAN_KYBER768_CLEAN_CRYPTO_CIPHERTEXTBYTES 1088
+#define PQCLEAN_KYBER768_CLEAN_CRYPTO_BYTES           32
 
 #define PQCLEAN_KYBER768_CLEAN_CRYPTO_ALGNAME "Kyber768"
 
-int PQCLEAN_KYBER768_CLEAN_crypto_kem_keypair(uint8_t *pk, uint8_t *sk);
+int PQCLEAN_KYBER768_CLEAN_crypto_kem_keypair(unsigned char *pk, unsigned char *sk);
 
-int PQCLEAN_KYBER768_CLEAN_crypto_kem_enc(uint8_t *ct, uint8_t *ss, const uint8_t *pk);
+int PQCLEAN_KYBER768_CLEAN_crypto_kem_enc(unsigned char *ct, unsigned char *ss, const unsigned char *pk);
+
+int PQCLEAN_KYBER768_CLEAN_crypto_kem_dec(unsigned char *ss, const unsigned char *ct, const unsigned char *sk);
 
-int PQCLEAN_KYBER768_CLEAN_crypto_kem_dec(uint8_t *ss, const uint8_t *ct, const uint8_t *sk);
 
 #endif
diff --git a/crypto_kem/kyber768/clean/cbd.c b/crypto_kem/kyber768/clean/cbd.c
index 15b46021..eac6a514 100644
--- a/crypto_kem/kyber768/clean/cbd.c
+++ b/crypto_kem/kyber768/clean/cbd.c
@@ -1,59 +1,53 @@
 #include "cbd.h"
 
+#include "params.h"
+
+#include <stdint.h>
+
 /*************************************************
- * Name:        load_littleendian
- *
- * Description: load bytes into a 64-bit integer
- *              in little-endian order
- *
- * Arguments:   - const unsigned char *x: pointer to input byte array
- *              - bytes:                  number of bytes to load, has to be <=
- *8
- *
- * Returns 64-bit unsigned integer loaded from x
- **************************************************/
-static uint64_t load_littleendian(const unsigned char *x, int bytes) {
-    int i;
-    uint64_t r = x[0];
-    for (i = 1; i < bytes; i++) {
-        r |= (uint64_t)x[i] << (8 * i);
-    }
+* Name:        load32_littleendian
+*
+* Description: load bytes into a 32-bit integer
+*              in little-endian order
+*
+* Arguments:   - const unsigned char *x: pointer to input byte array
+*
+* Returns 32-bit unsigned integer loaded from x
+**************************************************/
+static uint32_t load32_littleendian(const unsigned char *x) {
+    uint32_t r;
+    r  = (uint32_t)x[0];
+    r |= (uint32_t)x[1] << 8;
+    r |= (uint32_t)x[2] << 16;
+    r |= (uint32_t)x[3] << 24;
     return r;
 }
 
 /*************************************************
- * Name:        cbd
- *
- * Description: Given an array of uniformly random bytes, compute
- *              polynomial with coefficients distributed according to
- *              a centered binomial distribution with parameter KYBER_ETA
- *
- * Arguments:   - poly *r:                  pointer to output polynomial
- *              - const unsigned char *buf: pointer to input byte array
- **************************************************/
+* Name:        cbd
+*
+* Description: Given an array of uniformly random bytes, compute
+*              polynomial with coefficients distributed according to
+*              a centered binomial distribution with parameter KYBER_ETA
+*              specialized for KYBER_ETA=2
+*
+* Arguments:   - poly *r:                  pointer to output polynomial
+*              - const unsigned char *buf: pointer to input byte array
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_cbd(poly *r, const unsigned char *buf) {
-    uint32_t t, d, a[4], b[4];
+    uint32_t d, t;
+    int16_t a, b;
     int i, j;
 
-    for (i = 0; i < KYBER_N / 4; i++) {
-        t = (uint32_t)load_littleendian(buf + 4 * i, 4);
-        d = 0;
-        for (j = 0; j < 4; j++) {
-            d += (t >> j) & 0x11111111;
+    for (i = 0; i < KYBER_N / 8; i++) {
+        t = load32_littleendian(buf + 4 * i);
+        d  = t & 0x55555555;
+        d += (t >> 1) & 0x55555555;
+
+        for (j = 0; j < 8; j++) {
+            a = (d >>  4 * j)    & 0x3;
+            b = (d >> (4 * j + 2)) & 0x3;
+            r->coeffs[8 * i + j] = a - b;
         }
-
-        a[0] =  d        & 0xf;
-        b[0] = (d >>  4) & 0xf;
-        a[1] = (d >>  8) & 0xf;
-        b[1] = (d >> 12) & 0xf;
-        a[2] = (d >> 16) & 0xf;
-        b[2] = (d >> 20) & 0xf;
-        a[3] = (d >> 24) & 0xf;
-        b[3] = (d >> 28);
-
-        r->coeffs[4 * i + 0] = (uint16_t)(a[0] + KYBER_Q - b[0]);
-        r->coeffs[4 * i + 1] = (uint16_t)(a[1] + KYBER_Q - b[1]);
-        r->coeffs[4 * i + 2] = (uint16_t)(a[2] + KYBER_Q - b[2]);
-        r->coeffs[4 * i + 3] = (uint16_t)(a[3] + KYBER_Q - b[3]);
     }
 }
diff --git a/crypto_kem/kyber768/clean/cbd.h b/crypto_kem/kyber768/clean/cbd.h
index 88436fb0..1179a584 100644
--- a/crypto_kem/kyber768/clean/cbd.h
+++ b/crypto_kem/kyber768/clean/cbd.h
@@ -2,7 +2,6 @@
 #define CBD_H
 
 #include "poly.h"
-#include <stdint.h>
 
 void PQCLEAN_KYBER768_CLEAN_cbd(poly *r, const unsigned char *buf);
 
diff --git a/crypto_kem/kyber768/clean/indcpa.c b/crypto_kem/kyber768/clean/indcpa.c
index 1de2d599..3347434d 100644
--- a/crypto_kem/kyber768/clean/indcpa.c
+++ b/crypto_kem/kyber768/clean/indcpa.c
@@ -1,238 +1,239 @@
 #include "indcpa.h"
-#include "fips202.h"
+
 #include "ntt.h"
 #include "poly.h"
 #include "polyvec.h"
 #include "randombytes.h"
-#include <string.h>
+#include "symmetric.h"
+
+#include <stdint.h>
 
 /*************************************************
- * Name:        pack_pk
- *
- * Description: Serialize the public key as concatenation of the
- *              compressed and serialized vector of polynomials pk
- *              and the public seed used to generate the matrix A.
- *
- * Arguments:   unsigned char *r:          pointer to the output serialized
- *public key const poly *pk:            pointer to the input public-key
- *polynomial const unsigned char *seed: pointer to the input public seed
- **************************************************/
-static void pack_pk(unsigned char *r, const polyvec *pk, const unsigned char *seed) {
+* Name:        pack_pk
+*
+* Description: Serialize the public key as concatenation of the
+*              serialized vector of polynomials pk
+*              and the public seed used to generate the matrix A.
+*
+* Arguments:   unsigned char *r:          pointer to the output serialized public key
+*              const poly *pk:            pointer to the input public-key polynomial
+*              const unsigned char *seed: pointer to the input public seed
+**************************************************/
+static void pack_pk(unsigned char *r, polyvec *pk, const unsigned char *seed) {
     int i;
-    PQCLEAN_KYBER768_CLEAN_polyvec_compress(r, pk);
+    PQCLEAN_KYBER768_CLEAN_polyvec_tobytes(r, pk);
     for (i = 0; i < KYBER_SYMBYTES; i++) {
-        r[i + KYBER_POLYVECCOMPRESSEDBYTES] = seed[i];
+        r[i + KYBER_POLYVECBYTES] = seed[i];
     }
 }
 
 /*************************************************
- * Name:        unpack_pk
- *
- * Description: De-serialize and decompress public key from a byte array;
- *              approximate inverse of pack_pk
- *
- * Arguments:   - polyvec *pk:                   pointer to output public-key
- *vector of polynomials
- *              - unsigned char *seed:           pointer to output seed to
- *generate matrix A
- *              - const unsigned char *packedpk: pointer to input serialized
- *public key
- **************************************************/
+* Name:        unpack_pk
+*
+* Description: De-serialize public key from a byte array;
+*              approximate inverse of pack_pk
+*
+* Arguments:   - polyvec *pk:                   pointer to output public-key vector of polynomials
+*              - unsigned char *seed:           pointer to output seed to generate matrix A
+*              - const unsigned char *packedpk: pointer to input serialized public key
+**************************************************/
 static void unpack_pk(polyvec *pk, unsigned char *seed, const unsigned char *packedpk) {
     int i;
-    PQCLEAN_KYBER768_CLEAN_polyvec_decompress(pk, packedpk);
-
+    PQCLEAN_KYBER768_CLEAN_polyvec_frombytes(pk, packedpk);
     for (i = 0; i < KYBER_SYMBYTES; i++) {
-        seed[i] = packedpk[i + KYBER_POLYVECCOMPRESSEDBYTES];
+        seed[i] = packedpk[i + KYBER_POLYVECBYTES];
     }
 }
 
 /*************************************************
- * Name:        pack_ciphertext
- *
- * Description: Serialize the ciphertext as concatenation of the
- *              compressed and serialized vector of polynomials b
- *              and the compressed and serialized polynomial v
- *
- * Arguments:   unsigned char *r:          pointer to the output serialized
- *ciphertext const poly *pk:            pointer to the input vector of
- *polynomials b const unsigned char *seed: pointer to the input polynomial v
- **************************************************/
-static void pack_ciphertext(unsigned char *r, const polyvec *b, const poly *v) {
+* Name:        pack_sk
+*
+* Description: Serialize the secret key
+*
+* Arguments:   - unsigned char *r:  pointer to output serialized secret key
+*              - const polyvec *sk: pointer to input vector of polynomials (secret key)
+**************************************************/
+static void pack_sk(unsigned char *r, polyvec *sk) {
+    PQCLEAN_KYBER768_CLEAN_polyvec_tobytes(r, sk);
+}
+
+/*************************************************
+* Name:        unpack_sk
+*
+* Description: De-serialize the secret key;
+*              inverse of pack_sk
+*
+* Arguments:   - polyvec *sk:                   pointer to output vector of polynomials (secret key)
+*              - const unsigned char *packedsk: pointer to input serialized secret key
+**************************************************/
+static void unpack_sk(polyvec *sk, const unsigned char *packedsk) {
+    PQCLEAN_KYBER768_CLEAN_polyvec_frombytes(sk, packedsk);
+}
+
+/*************************************************
+* Name:        pack_ciphertext
+*
+* Description: Serialize the ciphertext as concatenation of the
+*              compressed and serialized vector of polynomials b
+*              and the compressed and serialized polynomial v
+*
+* Arguments:   unsigned char *r:          pointer to the output serialized ciphertext
+*              const poly *pk:            pointer to the input vector of polynomials b
+*              const unsigned char *seed: pointer to the input polynomial v
+**************************************************/
+static void pack_ciphertext(unsigned char *r, polyvec *b, poly *v) {
     PQCLEAN_KYBER768_CLEAN_polyvec_compress(r, b);
     PQCLEAN_KYBER768_CLEAN_poly_compress(r + KYBER_POLYVECCOMPRESSEDBYTES, v);
 }
 
 /*************************************************
- * Name:        unpack_ciphertext
- *
- * Description: De-serialize and decompress ciphertext from a byte array;
- *              approximate inverse of pack_ciphertext
- *
- * Arguments:   - polyvec *b:             pointer to the output vector of
- *polynomials b
- *              - poly *v:                pointer to the output polynomial v
- *              - const unsigned char *c: pointer to the input serialized
- *ciphertext
- **************************************************/
+* Name:        unpack_ciphertext
+*
+* Description: De-serialize and decompress ciphertext from a byte array;
+*              approximate inverse of pack_ciphertext
+*
+* Arguments:   - polyvec *b:             pointer to the output vector of polynomials b
+*              - poly *v:                pointer to the output polynomial v
+*              - const unsigned char *c: pointer to the input serialized ciphertext
+**************************************************/
 static void unpack_ciphertext(polyvec *b, poly *v, const unsigned char *c) {
     PQCLEAN_KYBER768_CLEAN_polyvec_decompress(b, c);
     PQCLEAN_KYBER768_CLEAN_poly_decompress(v, c + KYBER_POLYVECCOMPRESSEDBYTES);
 }
 
 /*************************************************
- * Name:        pack_sk
- *
- * Description: Serialize the secret key
- *
- * Arguments:   - unsigned char *r:  pointer to output serialized secret key
- *              - const polyvec *sk: pointer to input vector of polynomials
- *(secret key)
- **************************************************/
-static void pack_sk(unsigned char *r, const polyvec *sk) {
-    PQCLEAN_KYBER768_CLEAN_polyvec_tobytes(r, sk);
-}
-
-/*************************************************
- * Name:        unpack_sk
- *
- * Description: De-serialize the secret key;
- *              inverse of pack_sk
- *
- * Arguments:   - polyvec *sk:                   pointer to output vector of
- *polynomials (secret key)
- *              - const unsigned char *packedsk: pointer to input serialized
- *secret key
- **************************************************/
-static void unpack_sk(polyvec *sk, const unsigned char *packedsk) {
-    PQCLEAN_KYBER768_CLEAN_polyvec_frombytes(sk, packedsk);
-}
-
-#define gen_a(A, B) gen_matrix(A, B, 0)
-#define gen_at(A, B) gen_matrix(A, B, 1)
-
-/*************************************************
- * Name:        gen_matrix
- *
- * Description: Deterministically generate matrix A (or the transpose of A)
- *              from a seed. Entries of the matrix are polynomials that look
- *              uniformly random. Performs rejection sampling on output of
- *              SHAKE-128
- *
- * Arguments:   - polyvec *a:                pointer to ouptput matrix A
- *              - const unsigned char *seed: pointer to input seed
- *              - int transposed:            boolean deciding whether A or A^T
- *is generated
- **************************************************/
-static void gen_matrix(polyvec *a, const unsigned char *seed, int transposed) {
-    unsigned int pos = 0, ctr;
+* Name:        rej_uniform
+*
+* Description: Run rejection sampling on uniform random bytes to generate
+*              uniform random integers mod q
+*
+* Arguments:   - int16_t *r:               pointer to output buffer
+*              - unsigned int len:         requested number of 16-bit integers (uniform mod q)
+*              - const unsigned char *buf: pointer to input buffer (assumed to be uniform random bytes)
+*              - unsigned int buflen:      length of input buffer in bytes
+*
+* Returns number of sampled 16-bit integers (at most len)
+**************************************************/
+static unsigned int rej_uniform(int16_t *r, unsigned int len, const unsigned char *buf, unsigned int buflen) {
+    unsigned int ctr, pos;
     uint16_t val;
-    unsigned int nblocks;
-    const unsigned int maxnblocks = 4;
-    uint8_t buf[SHAKE128_RATE * /* maxnblocks = */ 4];
-    int i, j;
-    uint64_t state[25]; // SHAKE state
-    unsigned char extseed[KYBER_SYMBYTES + 2];
 
-    for (i = 0; i < KYBER_SYMBYTES; i++) {
-        extseed[i] = seed[i];
+    ctr = pos = 0;
+    while (ctr < len && pos + 2 <= buflen) {
+        val = buf[pos] | ((uint16_t)buf[pos + 1] << 8);
+        pos += 2;
+
+        if (val < 19 * KYBER_Q) {
+            val -= (val >> 12) * KYBER_Q; // Barrett reduction
+            r[ctr++] = (int16_t)val;
+        }
     }
 
+    return ctr;
+}
+
+#define gen_a(A,B)  PQCLEAN_KYBER768_CLEAN_gen_matrix(A,B,0)
+#define gen_at(A,B) PQCLEAN_KYBER768_CLEAN_gen_matrix(A,B,1)
+
+#define GENMATRIX_MAXNBLOCKS ((530 + XOF_BLOCKBYTES) / XOF_BLOCKBYTES) /* 530 is expected number of required bytes */
+
+
+/*************************************************
+* Name:        gen_matrix
+*
+* Description: Deterministically generate matrix A (or the transpose of A)
+*              from a seed. Entries of the matrix are polynomials that look
+*              uniformly random. Performs rejection sampling on output of
+*              a XOF
+*
+* Arguments:   - polyvec *a:                pointer to ouptput matrix A
+*              - const unsigned char *seed: pointer to input seed
+*              - int transposed:            boolean deciding whether A or A^T is generated
+**************************************************/
+static void PQCLEAN_KYBER768_CLEAN_gen_matrix(polyvec *a, const unsigned char *seed, int transposed) {
+    unsigned int ctr;
+    unsigned char i, j;
+    unsigned char buf[XOF_BLOCKBYTES * GENMATRIX_MAXNBLOCKS + 1];
+    xof_state state;
+
     for (i = 0; i < KYBER_K; i++) {
         for (j = 0; j < KYBER_K; j++) {
-            ctr = pos = 0;
-            nblocks = maxnblocks;
             if (transposed) {
-                extseed[KYBER_SYMBYTES] = (unsigned char)i;
-                extseed[KYBER_SYMBYTES + 1] = (unsigned char)j;
+                xof_absorb(&state, seed, i, j);
             } else {
-                extseed[KYBER_SYMBYTES] = (unsigned char)j;
-                extseed[KYBER_SYMBYTES + 1] = (unsigned char)i;
+                xof_absorb(&state, seed, j, i);
             }
 
-            shake128_absorb(state, extseed, KYBER_SYMBYTES + 2);
-            shake128_squeezeblocks(buf, nblocks, state);
+            xof_squeezeblocks(buf, GENMATRIX_MAXNBLOCKS, &state);
+            ctr = rej_uniform(a[i].vec[j].coeffs, KYBER_N, buf, GENMATRIX_MAXNBLOCKS * XOF_BLOCKBYTES);
 
             while (ctr < KYBER_N) {
-                val = (buf[pos] | ((uint16_t)buf[pos + 1] << 8)) & 0x1fff;
-                if (val < KYBER_Q) {
-                    a[i].vec[j].coeffs[ctr++] = val;
-                }
-                pos += 2;
-
-                if (pos > SHAKE128_RATE * nblocks - 2) {
-                    nblocks = 1;
-                    shake128_squeezeblocks(buf, nblocks, state);
-                    pos = 0;
-                }
+                xof_squeezeblocks(buf, 1, &state);
+                ctr += rej_uniform(a[i].vec[j].coeffs + ctr, KYBER_N - ctr, buf, XOF_BLOCKBYTES);
             }
         }
     }
 }
 
 /*************************************************
- * Name:        indcpa_keypair
- *
- * Description: Generates public and private key for the CPA-secure
- *              public-key encryption scheme underlying Kyber
- *
- * Arguments:   - unsigned char *pk: pointer to output public key (of length
- *KYBER_INDCPA_PUBLICKEYBYTES bytes)
- *              - unsigned char *sk: pointer to output private key (of length
- *KYBER_INDCPA_SECRETKEYBYTES bytes)
- **************************************************/
+* Name:        indcpa_keypair
+*
+* Description: Generates public and private key for the CPA-secure
+*              public-key encryption scheme underlying Kyber
+*
+* Arguments:   - unsigned char *pk: pointer to output public key (of length KYBER_INDCPA_PUBLICKEYBYTES bytes)
+*              - unsigned char *sk: pointer to output private key (of length KYBER_INDCPA_SECRETKEYBYTES bytes)
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_indcpa_keypair(unsigned char *pk, unsigned char *sk) {
     polyvec a[KYBER_K], e, pkpv, skpv;
-    unsigned char buf[KYBER_SYMBYTES + KYBER_SYMBYTES];
+    unsigned char buf[2 * KYBER_SYMBYTES];
     unsigned char *publicseed = buf;
     unsigned char *noiseseed = buf + KYBER_SYMBYTES;
     int i;
     unsigned char nonce = 0;
 
     randombytes(buf, KYBER_SYMBYTES);
-    sha3_512(buf, buf, KYBER_SYMBYTES);
+    hash_g(buf, buf, KYBER_SYMBYTES);
 
     gen_a(a, publicseed);
 
     for (i = 0; i < KYBER_K; i++) {
         PQCLEAN_KYBER768_CLEAN_poly_getnoise(skpv.vec + i, noiseseed, nonce++);
     }
-
-    PQCLEAN_KYBER768_CLEAN_polyvec_ntt(&skpv);
-
     for (i = 0; i < KYBER_K; i++) {
         PQCLEAN_KYBER768_CLEAN_poly_getnoise(e.vec + i, noiseseed, nonce++);
     }
 
+    PQCLEAN_KYBER768_CLEAN_polyvec_ntt(&skpv);
+    PQCLEAN_KYBER768_CLEAN_polyvec_ntt(&e);
+
     // matrix-vector multiplication
     for (i = 0; i < KYBER_K; i++) {
-        PQCLEAN_KYBER768_CLEAN_polyvec_pointwise_acc(&pkpv.vec[i], &skpv, a + i);
+        PQCLEAN_KYBER768_CLEAN_polyvec_pointwise_acc(&pkpv.vec[i], &a[i], &skpv);
+        PQCLEAN_KYBER768_CLEAN_poly_frommont(&pkpv.vec[i]);
     }
 
-    PQCLEAN_KYBER768_CLEAN_polyvec_invntt(&pkpv);
     PQCLEAN_KYBER768_CLEAN_polyvec_add(&pkpv, &pkpv, &e);
+    PQCLEAN_KYBER768_CLEAN_polyvec_reduce(&pkpv);
 
     pack_sk(sk, &skpv);
     pack_pk(pk, &pkpv, publicseed);
 }
 
 /*************************************************
- * Name:        indcpa_enc
- *
- * Description: Encryption function of the CPA-secure
- *              public-key encryption scheme underlying Kyber.
- *
- * Arguments:   - unsigned char *c:          pointer to output ciphertext (of
- *length KYBER_INDCPA_BYTES bytes)
- *              - const unsigned char *m:    pointer to input message (of length
- *KYBER_INDCPA_MSGBYTES bytes)
- *              - const unsigned char *pk:   pointer to input public key (of
- *length KYBER_INDCPA_PUBLICKEYBYTES bytes)
- *              - const unsigned char *coin: pointer to input random coins used
- *as seed (of length KYBER_SYMBYTES bytes) to deterministically generate all
- *randomness
- **************************************************/
-void PQCLEAN_KYBER768_CLEAN_indcpa_enc(unsigned char *c, const unsigned char *m,
+* Name:        indcpa_enc
+*
+* Description: Encryption function of the CPA-secure
+*              public-key encryption scheme underlying Kyber.
+*
+* Arguments:   - unsigned char *c:          pointer to output ciphertext (of length KYBER_INDCPA_BYTES bytes)
+*              - const unsigned char *m:    pointer to input message (of length KYBER_INDCPA_MSGBYTES bytes)
+*              - const unsigned char *pk:   pointer to input public key (of length KYBER_INDCPA_PUBLICKEYBYTES bytes)
+*              - const unsigned char *coin: pointer to input random coins used as seed (of length KYBER_SYMBYTES bytes)
+*                                           to deterministically generate all randomness
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_indcpa_enc(unsigned char *c,
+                                       const unsigned char *m,
                                        const unsigned char *pk,
                                        const unsigned char *coins) {
     polyvec sp, pkpv, ep, at[KYBER_K], bp;
@@ -242,56 +243,50 @@ void PQCLEAN_KYBER768_CLEAN_indcpa_enc(unsigned char *c, const unsigned char *m,
     unsigned char nonce = 0;
 
     unpack_pk(&pkpv, seed, pk);
-
     PQCLEAN_KYBER768_CLEAN_poly_frommsg(&k, m);
-
-    PQCLEAN_KYBER768_CLEAN_polyvec_ntt(&pkpv);
-
     gen_at(at, seed);
 
     for (i = 0; i < KYBER_K; i++) {
         PQCLEAN_KYBER768_CLEAN_poly_getnoise(sp.vec + i, coins, nonce++);
     }
-
-    PQCLEAN_KYBER768_CLEAN_polyvec_ntt(&sp);
-
     for (i = 0; i < KYBER_K; i++) {
         PQCLEAN_KYBER768_CLEAN_poly_getnoise(ep.vec + i, coins, nonce++);
     }
+    PQCLEAN_KYBER768_CLEAN_poly_getnoise(&epp, coins, nonce++);
+
+    PQCLEAN_KYBER768_CLEAN_polyvec_ntt(&sp);
 
     // matrix-vector multiplication
     for (i = 0; i < KYBER_K; i++) {
-        PQCLEAN_KYBER768_CLEAN_polyvec_pointwise_acc(&bp.vec[i], &sp, at + i);
+        PQCLEAN_KYBER768_CLEAN_polyvec_pointwise_acc(&bp.vec[i], &at[i], &sp);
     }
 
-    PQCLEAN_KYBER768_CLEAN_polyvec_invntt(&bp);
-    PQCLEAN_KYBER768_CLEAN_polyvec_add(&bp, &bp, &ep);
-
     PQCLEAN_KYBER768_CLEAN_polyvec_pointwise_acc(&v, &pkpv, &sp);
+
+    PQCLEAN_KYBER768_CLEAN_polyvec_invntt(&bp);
     PQCLEAN_KYBER768_CLEAN_poly_invntt(&v);
 
-    PQCLEAN_KYBER768_CLEAN_poly_getnoise(&epp, coins, nonce++);
-
+    PQCLEAN_KYBER768_CLEAN_polyvec_add(&bp, &bp, &ep);
     PQCLEAN_KYBER768_CLEAN_poly_add(&v, &v, &epp);
     PQCLEAN_KYBER768_CLEAN_poly_add(&v, &v, &k);
+    PQCLEAN_KYBER768_CLEAN_polyvec_reduce(&bp);
+    PQCLEAN_KYBER768_CLEAN_poly_reduce(&v);
 
     pack_ciphertext(c, &bp, &v);
 }
 
 /*************************************************
- * Name:        indcpa_dec
- *
- * Description: Decryption function of the CPA-secure
- *              public-key encryption scheme underlying Kyber.
- *
- * Arguments:   - unsigned char *m:        pointer to output decrypted message
- *(of length KYBER_INDCPA_MSGBYTES)
- *              - const unsigned char *c:  pointer to input ciphertext (of
- *length KYBER_INDCPA_BYTES)
- *              - const unsigned char *sk: pointer to input secret key (of
- *length KYBER_INDCPA_SECRETKEYBYTES)
- **************************************************/
-void PQCLEAN_KYBER768_CLEAN_indcpa_dec(unsigned char *m, const unsigned char *c,
+* Name:        indcpa_dec
+*
+* Description: Decryption function of the CPA-secure
+*              public-key encryption scheme underlying Kyber.
+*
+* Arguments:   - unsigned char *m:        pointer to output decrypted message (of length KYBER_INDCPA_MSGBYTES)
+*              - const unsigned char *c:  pointer to input ciphertext (of length KYBER_INDCPA_BYTES)
+*              - const unsigned char *sk: pointer to input secret key (of length KYBER_INDCPA_SECRETKEYBYTES)
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_indcpa_dec(unsigned char *m,
+                                       const unsigned char *c,
                                        const unsigned char *sk) {
     polyvec bp, skpv;
     poly v, mp;
@@ -300,11 +295,11 @@ void PQCLEAN_KYBER768_CLEAN_indcpa_dec(unsigned char *m, const unsigned char *c,
     unpack_sk(&skpv, sk);
 
     PQCLEAN_KYBER768_CLEAN_polyvec_ntt(&bp);
-
     PQCLEAN_KYBER768_CLEAN_polyvec_pointwise_acc(&mp, &skpv, &bp);
     PQCLEAN_KYBER768_CLEAN_poly_invntt(&mp);
 
-    PQCLEAN_KYBER768_CLEAN_poly_sub(&mp, &mp, &v);
+    PQCLEAN_KYBER768_CLEAN_poly_sub(&mp, &v, &mp);
+    PQCLEAN_KYBER768_CLEAN_poly_reduce(&mp);
 
     PQCLEAN_KYBER768_CLEAN_poly_tomsg(m, &mp);
 }
diff --git a/crypto_kem/kyber768/clean/kem.c b/crypto_kem/kyber768/clean/kem.c
index bda53edf..aada88ff 100644
--- a/crypto_kem/kyber768/clean/kem.c
+++ b/crypto_kem/kyber768/clean/kem.c
@@ -1,108 +1,102 @@
 #include "api.h"
-#include "fips202.h"
 #include "indcpa.h"
 #include "params.h"
 #include "randombytes.h"
+#include "symmetric.h"
 #include "verify.h"
 
+#include <stdlib.h>
+
+#include <stdlib.h>
+
 /*************************************************
- * Name:        crypto_kem_keypair
- *
- * Description: Generates public and private key
- *              for CCA-secure Kyber key encapsulation mechanism
- *
- * Arguments:   - unsigned char *pk: pointer to output public key (an already
- *allocated array of CRYPTO_PUBLICKEYBYTES bytes)
- *              - unsigned char *sk: pointer to output private key (an already
- *allocated array of CRYPTO_SECRETKEYBYTES bytes)
- *
- * Returns 0 (success)
- **************************************************/
+* Name:        crypto_kem_keypair
+*
+* Description: Generates public and private key
+*              for CCA-secure Kyber key encapsulation mechanism
+*
+* Arguments:   - unsigned char *pk: pointer to output public key (an already allocated array of CRYPTO_PUBLICKEYBYTES bytes)
+*              - unsigned char *sk: pointer to output private key (an already allocated array of CRYPTO_SECRETKEYBYTES bytes)
+*
+* Returns 0 (success)
+**************************************************/
 int PQCLEAN_KYBER768_CLEAN_crypto_kem_keypair(unsigned char *pk, unsigned char *sk) {
     size_t i;
     PQCLEAN_KYBER768_CLEAN_indcpa_keypair(pk, sk);
     for (i = 0; i < KYBER_INDCPA_PUBLICKEYBYTES; i++) {
         sk[i + KYBER_INDCPA_SECRETKEYBYTES] = pk[i];
     }
-    sha3_256(sk + KYBER_SECRETKEYBYTES - 2 * KYBER_SYMBYTES, pk, KYBER_PUBLICKEYBYTES);
-    randombytes(sk + KYBER_SECRETKEYBYTES - KYBER_SYMBYTES, KYBER_SYMBYTES);             /* Value z for pseudo-random output on reject */
+    hash_h(sk + KYBER_SECRETKEYBYTES - 2 * KYBER_SYMBYTES, pk, KYBER_PUBLICKEYBYTES);
+    randombytes(sk + KYBER_SECRETKEYBYTES - KYBER_SYMBYTES, KYBER_SYMBYTES);    /* Value z for pseudo-random output on reject */
     return 0;
 }
 
 /*************************************************
- * Name:        crypto_kem_enc
- *
- * Description: Generates cipher text and shared
- *              secret for given public key
- *
- * Arguments:   - unsigned char *ct:       pointer to output cipher text (an
- *already allocated array of CRYPTO_CIPHERTEXTBYTES bytes)
- *              - unsigned char *ss:       pointer to output shared secret (an
- *already allocated array of CRYPTO_BYTES bytes)
- *              - const unsigned char *pk: pointer to input public key (an
- *already allocated array of CRYPTO_PUBLICKEYBYTES bytes)
- *
- * Returns 0 (success)
- **************************************************/
+* Name:        crypto_kem_enc
+*
+* Description: Generates cipher text and shared
+*              secret for given public key
+*
+* Arguments:   - unsigned char *ct:       pointer to output cipher text (an already allocated array of CRYPTO_CIPHERTEXTBYTES bytes)
+*              - unsigned char *ss:       pointer to output shared secret (an already allocated array of CRYPTO_BYTES bytes)
+*              - const unsigned char *pk: pointer to input public key (an already allocated array of CRYPTO_PUBLICKEYBYTES bytes)
+*
+* Returns 0 (success)
+**************************************************/
 int PQCLEAN_KYBER768_CLEAN_crypto_kem_enc(unsigned char *ct, unsigned char *ss, const unsigned char *pk) {
-    unsigned char kr[2 * KYBER_SYMBYTES]; /* Will contain key, coins */
+    unsigned char  kr[2 * KYBER_SYMBYTES];                                   /* Will contain key, coins */
     unsigned char buf[2 * KYBER_SYMBYTES];
 
     randombytes(buf, KYBER_SYMBYTES);
-    sha3_256(buf, buf, KYBER_SYMBYTES);                            /* Don't release system RNG output */
+    hash_h(buf, buf, KYBER_SYMBYTES);                                        /* Don't release system RNG output */
 
-    sha3_256(buf + KYBER_SYMBYTES, pk, KYBER_PUBLICKEYBYTES);      /* Multitarget countermeasure for coins + contributory KEM */
-    sha3_512(kr, buf, 2 * KYBER_SYMBYTES);
+    hash_h(buf + KYBER_SYMBYTES, pk, KYBER_PUBLICKEYBYTES);                  /* Multitarget countermeasure for coins + contributory KEM */
+    hash_g(kr, buf, 2 * KYBER_SYMBYTES);
 
-    PQCLEAN_KYBER768_CLEAN_indcpa_enc(ct, buf, pk, kr + KYBER_SYMBYTES); /* coins are in kr+KYBER_SYMBYTES */
+    PQCLEAN_KYBER768_CLEAN_indcpa_enc(ct, buf, pk, kr + KYBER_SYMBYTES);                            /* coins are in kr+KYBER_SYMBYTES */
 
-    sha3_256(kr + KYBER_SYMBYTES, ct, KYBER_CIPHERTEXTBYTES);      /* overwrite coins in kr with H(c) */
-    sha3_256(ss, kr, 2 * KYBER_SYMBYTES);                          /* hash concatenation of pre-k and H(c) to k */
+    hash_h(kr + KYBER_SYMBYTES, ct, KYBER_CIPHERTEXTBYTES);                  /* overwrite coins in kr with H(c) */
+    kdf(ss, kr, 2 * KYBER_SYMBYTES);                                         /* hash concatenation of pre-k and H(c) to k */
     return 0;
 }
 
 /*************************************************
- * Name:        crypto_kem_dec
- *
- * Description: Generates shared secret for given
- *              cipher text and private key
- *
- * Arguments:   - unsigned char *ss:       pointer to output shared secret (an
- *already allocated array of CRYPTO_BYTES bytes)
- *              - const unsigned char *ct: pointer to input cipher text (an
- *already allocated array of CRYPTO_CIPHERTEXTBYTES bytes)
- *              - const unsigned char *sk: pointer to input private key (an
- *already allocated array of CRYPTO_SECRETKEYBYTES bytes)
- *
- * Returns 0.
- *
- * On failure, ss will contain a pseudo-random value.
- **************************************************/
+* Name:        crypto_kem_dec
+*
+* Description: Generates shared secret for given
+*              cipher text and private key
+*
+* Arguments:   - unsigned char *ss:       pointer to output shared secret (an already allocated array of CRYPTO_BYTES bytes)
+*              - const unsigned char *ct: pointer to input cipher text (an already allocated array of CRYPTO_CIPHERTEXTBYTES bytes)
+*              - const unsigned char *sk: pointer to input private key (an already allocated array of CRYPTO_SECRETKEYBYTES bytes)
+*
+* Returns 0.
+*
+* On failure, ss will contain a pseudo-random value.
+**************************************************/
 int PQCLEAN_KYBER768_CLEAN_crypto_kem_dec(unsigned char *ss, const unsigned char *ct, const unsigned char *sk) {
     size_t i;
-    int fail;
+    unsigned char fail;
     unsigned char cmp[KYBER_CIPHERTEXTBYTES];
     unsigned char buf[2 * KYBER_SYMBYTES];
-    unsigned char
-    kr[2 * KYBER_SYMBYTES];                                                                                     /* Will contain key, coins, qrom-hash */
+    unsigned char kr[2 * KYBER_SYMBYTES];                                    /* Will contain key, coins */
     const unsigned char *pk = sk + KYBER_INDCPA_SECRETKEYBYTES;
 
     PQCLEAN_KYBER768_CLEAN_indcpa_dec(buf, ct, sk);
 
-    for (i = 0; i < KYBER_SYMBYTES; i++) {                                                                      /* Multitarget countermeasure for coins + contributory KEM */
-        buf[KYBER_SYMBYTES + i] = sk[KYBER_SECRETKEYBYTES - 2 * KYBER_SYMBYTES + i];                            /* Save hash by storing H(pk) in sk */
+    for (i = 0; i < KYBER_SYMBYTES; i++) {                                   /* Multitarget countermeasure for coins + contributory KEM */
+        buf[KYBER_SYMBYTES + i] = sk[KYBER_SECRETKEYBYTES - 2 * KYBER_SYMBYTES + i];    /* Save hash by storing H(pk) in sk */
     }
-    sha3_512(kr, buf, 2 * KYBER_SYMBYTES);
+    hash_g(kr, buf, 2 * KYBER_SYMBYTES);
 
-    PQCLEAN_KYBER768_CLEAN_indcpa_enc(cmp, buf, pk, kr + KYBER_SYMBYTES);                                             /* coins are in kr+KYBER_SYMBYTES */
+    PQCLEAN_KYBER768_CLEAN_indcpa_enc(cmp, buf, pk, kr + KYBER_SYMBYTES);                           /* coins are in kr+KYBER_SYMBYTES */
 
     fail = PQCLEAN_KYBER768_CLEAN_verify(ct, cmp, KYBER_CIPHERTEXTBYTES);
 
-    sha3_256(kr + KYBER_SYMBYTES, ct, KYBER_CIPHERTEXTBYTES);                                                   /* overwrite coins in kr with H(c)  */
+    hash_h(kr + KYBER_SYMBYTES, ct, KYBER_CIPHERTEXTBYTES);                  /* overwrite coins in kr with H(c)  */
 
-    PQCLEAN_KYBER768_CLEAN_cmov(kr, sk + KYBER_SECRETKEYBYTES - KYBER_SYMBYTES, KYBER_SYMBYTES, (unsigned char)fail); /* Overwrite pre-k with z on re-encryption failure */
-
-    sha3_256(ss, kr, 2 * KYBER_SYMBYTES);                                                                       /* hash concatenation of pre-k and H(c) to k */
+    PQCLEAN_KYBER768_CLEAN_cmov(kr, sk + KYBER_SECRETKEYBYTES - KYBER_SYMBYTES, KYBER_SYMBYTES, fail); /* Overwrite pre-k with z on re-encryption failure */
 
+    kdf(ss, kr, 2 * KYBER_SYMBYTES);                                         /* hash concatenation of pre-k and H(c) to k */
     return 0;
 }
diff --git a/crypto_kem/kyber768/clean/ntt.c b/crypto_kem/kyber768/clean/ntt.c
index 53fc4595..ea98c4c6 100644
--- a/crypto_kem/kyber768/clean/ntt.c
+++ b/crypto_kem/kyber768/clean/ntt.c
@@ -1,80 +1,157 @@
 #include "ntt.h"
-#include "inttypes.h"
+
 #include "params.h"
 #include "reduce.h"
 
-extern const uint16_t PQCLEAN_KYBER768_CLEAN_omegas_inv_bitrev_montgomery[];
-extern const uint16_t PQCLEAN_KYBER768_CLEAN_psis_inv_montgomery[];
-extern const uint16_t PQCLEAN_KYBER768_CLEAN_zetas[];
+#include <stdint.h>
+
+/* Code to generate zetas and zetas_inv used in the number-theoretic transform:
+
+#define KYBER_ROOT_OF_UNITY 17
+
+static const uint16_t tree[128] = {
+  0, 64, 32, 96, 16, 80, 48, 112, 8, 72, 40, 104, 24, 88, 56, 120,
+  4, 68, 36, 100, 20, 84, 52, 116, 12, 76, 44, 108, 28, 92, 60, 124,
+  2, 66, 34, 98, 18, 82, 50, 114, 10, 74, 42, 106, 26, 90, 58, 122,
+  6, 70, 38, 102, 22, 86, 54, 118, 14, 78, 46, 110, 30, 94, 62, 126,
+  1, 65, 33, 97, 17, 81, 49, 113, 9, 73, 41, 105, 25, 89, 57, 121,
+  5, 69, 37, 101, 21, 85, 53, 117, 13, 77, 45, 109, 29, 93, 61, 125,
+  3, 67, 35, 99, 19, 83, 51, 115, 11, 75, 43, 107, 27, 91, 59, 123,
+  7, 71, 39, 103, 23, 87, 55, 119, 15, 79, 47, 111, 31, 95, 63, 127};
+
+
+static int16_t fqmul(int16_t a, int16_t b) {
+  return montgomery_reduce((int32_t)a*b);
+}
+
+void init_ntt() {
+  unsigned int i, j, k;
+  int16_t tmp[128];
+
+  tmp[0] = MONT;
+  for(i = 1; i < 128; ++i)
+    tmp[i] = fqmul(tmp[i-1], KYBER_ROOT_OF_UNITY*MONT % KYBER_Q);
+
+  for(i = 0; i < 128; ++i)
+    zetas[i] = tmp[tree[i]];
+
+  k = 0;
+  for(i = 64; i >= 1; i >>= 1)
+    for(j = i; j < 2*i; ++j)
+      zetas_inv[k++] = -tmp[128 - tree[j]];
+
+  zetas_inv[127] = MONT * (MONT * (KYBER_Q - 1) * ((KYBER_Q - 1)/128) % KYBER_Q) % KYBER_Q;
+}
+
+*/
+int16_t PQCLEAN_KYBER768_CLEAN_zetas[128] = {
+    2285, 2571, 2970, 1812, 1493, 1422, 287, 202, 3158, 622, 1577, 182, 962, 2127, 1855, 1468,
+    573, 2004, 264, 383, 2500, 1458, 1727, 3199, 2648, 1017, 732, 608, 1787, 411, 3124, 1758,
+    1223, 652, 2777, 1015, 2036, 1491, 3047, 1785, 516, 3321, 3009, 2663, 1711, 2167, 126, 1469,
+    2476, 3239, 3058, 830, 107, 1908, 3082, 2378, 2931, 961, 1821, 2604, 448, 2264, 677, 2054,
+    2226, 430, 555, 843, 2078, 871, 1550, 105, 422, 587, 177, 3094, 3038, 2869, 1574, 1653,
+    3083, 778, 1159, 3182, 2552, 1483, 2727, 1119, 1739, 644, 2457, 349, 418, 329, 3173, 3254,
+    817, 1097, 603, 610, 1322, 2044, 1864, 384, 2114, 3193, 1218, 1994, 2455, 220, 2142, 1670,
+    2144, 1799, 2051, 794, 1819, 2475, 2459, 478, 3221, 3021, 996, 991, 958, 1869, 1522, 1628
+};
+
+int16_t PQCLEAN_KYBER768_CLEAN_zetas_inv[128] = {
+    1701, 1807, 1460, 2371, 2338, 2333, 308, 108, 2851, 870, 854, 1510, 2535, 1278, 1530, 1185,
+    1659, 1187, 3109, 874, 1335, 2111, 136, 1215, 2945, 1465, 1285, 2007, 2719, 2726, 2232, 2512,
+    75, 156, 3000, 2911, 2980, 872, 2685, 1590, 2210, 602, 1846, 777, 147, 2170, 2551, 246,
+    1676, 1755, 460, 291, 235, 3152, 2742, 2907, 3224, 1779, 2458, 1251, 2486, 2774, 2899, 1103,
+    1275, 2652, 1065, 2881, 725, 1508, 2368, 398, 951, 247, 1421, 3222, 2499, 271, 90, 853,
+    1860, 3203, 1162, 1618, 666, 320, 8, 2813, 1544, 282, 1838, 1293, 2314, 552, 2677, 2106,
+    1571, 205, 2918, 1542, 2721, 2597, 2312, 681, 130, 1602, 1871, 829, 2946, 3065, 1325, 2756,
+    1861, 1474, 1202, 2367, 3147, 1752, 2707, 171, 3127, 3042, 1907, 1836, 1517, 359, 758, 1441
+};
+
 
 /*************************************************
- * Name:        ntt
- *
- * Description: Computes negacyclic number-theoretic transform (NTT) of
- *              a polynomial (vector of 256 coefficients) in place;
- *              inputs assumed to be in normal order, output in bitreversed
- *order
- *
- * Arguments:   - uint16_t *p: pointer to in/output polynomial
- **************************************************/
-void PQCLEAN_KYBER768_CLEAN_ntt(uint16_t *p) {
-    int level, start, j, k;
-    uint16_t zeta, t;
+* Name:        fqmul
+*
+* Description: Multiplication followed by Montgomery reduction
+*
+* Arguments:   - int16_t a: first factor
+*              - int16_t b: second factor
+*
+* Returns 16-bit integer congruent to a*b*R^{-1} mod q
+**************************************************/
+static int16_t fqmul(int16_t a, int16_t b) {
+    return PQCLEAN_KYBER768_CLEAN_montgomery_reduce((int32_t)a * b);
+}
+
+/*************************************************
+* Name:        ntt
+*
+* Description: Inplace number-theoretic transform (NTT) in Rq
+*              input is in standard order, output is in bitreversed order
+*
+* Arguments:   - int16_t *poly: pointer to input/output vector of 256 elements of Zq
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_ntt(int16_t *poly) {
+    unsigned int len, start, j, k;
+    int16_t t, zeta;
 
     k = 1;
-    for (level = 7; level >= 0; level--) {
-        for (start = 0; start < KYBER_N; start = j + (1 << level)) {
+    for (len = 128; len >= 2; len >>= 1) {
+        for (start = 0; start < 256; start = j + len) {
             zeta = PQCLEAN_KYBER768_CLEAN_zetas[k++];
-            for (j = start; j < start + (1 << level); ++j) {
-                t = PQCLEAN_KYBER768_CLEAN_montgomery_reduce((uint32_t)zeta * p[j + (1 << level)]);
-
-                p[j + (1 << level)] = PQCLEAN_KYBER768_CLEAN_barrett_reduce(p[j] + 4 * KYBER_Q - t);
-
-                if (level & 1) {     /* odd level */
-                    p[j] = p[j] + t; /* Omit reduction (be lazy) */
-                } else {
-                    p[j] = PQCLEAN_KYBER768_CLEAN_barrett_reduce(p[j] + t);
-                }
+            for (j = start; j < start + len; ++j) {
+                t = fqmul(zeta, poly[j + len]);
+                poly[j + len] = poly[j] - t;
+                poly[j] = poly[j] + t;
             }
         }
     }
 }
 
 /*************************************************
- * Name:        invntt
- *
- * Description: Computes inverse of negacyclic number-theoretic transform (NTT)
- *of a polynomial (vector of 256 coefficients) in place; inputs assumed to be in
- *bitreversed order, output in normal order
- *
- * Arguments:   - uint16_t *a: pointer to in/output polynomial
- **************************************************/
-void PQCLEAN_KYBER768_CLEAN_invntt(uint16_t *a) {
-    int start, j, jTwiddle, level;
-    uint16_t temp, W;
-    uint32_t t;
+* Name:        invntt
+*
+* Description: Inplace inverse number-theoretic transform in Rq
+*              input is in bitreversed order, output is in standard order
+*
+* Arguments:   - int16_t *poly: pointer to input/output vector of 256 elements of Zq
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_invntt(int16_t *poly) {
+    unsigned int start, len, j, k;
+    int16_t t, zeta;
 
-    for (level = 0; level < 8; level++) {
-        for (start = 0; start < (1 << level); start++) {
-            jTwiddle = 0;
-            for (j = start; j < KYBER_N - 1; j += 2 * (1 << level)) {
-                W = PQCLEAN_KYBER768_CLEAN_omegas_inv_bitrev_montgomery[jTwiddle++];
-                temp = a[j];
-
-                if (level & 1) { /* odd level */
-                    a[j] = PQCLEAN_KYBER768_CLEAN_barrett_reduce((temp + a[j + (1 << level)]));
-                } else {
-                    a[j] = (temp + a[j + (1 << level)]); /* Omit reduction (be lazy) */
-                }
-
-                t = (W * ((uint32_t)temp + 4 * KYBER_Q - a[j + (1 << level)]));
-
-                a[j + (1 << level)] = PQCLEAN_KYBER768_CLEAN_montgomery_reduce(t);
+    k = 0;
+    for (len = 2; len <= 128; len <<= 1) {
+        for (start = 0; start < 256; start = j + len) {
+            zeta = PQCLEAN_KYBER768_CLEAN_zetas_inv[k++];
+            for (j = start; j < start + len; ++j) {
+                t = poly[j];
+                poly[j] = PQCLEAN_KYBER768_CLEAN_barrett_reduce(t + poly[j + len]);
+                poly[j + len] = t - poly[j + len];
+                poly[j + len] = fqmul(zeta, poly[j + len]);
             }
         }
     }
 
-    for (j = 0; j < KYBER_N; j++) {
-        a[j] = PQCLEAN_KYBER768_CLEAN_montgomery_reduce((a[j] * PQCLEAN_KYBER768_CLEAN_psis_inv_montgomery[j]));
+    for (j = 0; j < 256; ++j) {
+        poly[j] = fqmul(poly[j], PQCLEAN_KYBER768_CLEAN_zetas_inv[127]);
     }
 }
+
+/*************************************************
+* Name:        PQCLEAN_KYBER768_CLEAN_basemul
+*
+* Description: Multiplication of polynomials in Zq[X]/((X^2-zeta))
+*              used for multiplication of elements in Rq in NTT domain
+*
+* Arguments:   - int16_t r[2]: pointer to the output polynomial
+*              - const int16_t a[2]: pointer to the first factor
+*              - const int16_t b[2]: pointer to the second factor
+*              - int16_t zeta: integer defining the reduction polynomial
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_basemul(int16_t r[2], const int16_t a[2], const int16_t b[2], int16_t zeta) {
+    r[0]  = fqmul(a[1], b[1]);
+    r[0]  = fqmul(r[0], zeta);
+    r[0] += fqmul(a[0], b[0]);
+
+    r[1]  = fqmul(a[0], b[1]);
+    r[1] += fqmul(a[1], b[0]);
+}
diff --git a/crypto_kem/kyber768/clean/ntt.h b/crypto_kem/kyber768/clean/ntt.h
index e01d83a2..e0e169f1 100644
--- a/crypto_kem/kyber768/clean/ntt.h
+++ b/crypto_kem/kyber768/clean/ntt.h
@@ -3,7 +3,11 @@
 
 #include <stdint.h>
 
-void PQCLEAN_KYBER768_CLEAN_ntt(uint16_t *p);
-void PQCLEAN_KYBER768_CLEAN_invntt(uint16_t *a);
+extern int16_t PQCLEAN_KYBER768_CLEAN_zetas[128];
+extern int16_t PQCLEAN_KYBER768_CLEAN_zetas_inv[128];
+
+void PQCLEAN_KYBER768_CLEAN_ntt(int16_t *poly);
+void PQCLEAN_KYBER768_CLEAN_invntt(int16_t *poly);
+void PQCLEAN_KYBER768_CLEAN_basemul(int16_t r[2], const int16_t a[2], const int16_t b[2], int16_t zeta);
 
 #endif
diff --git a/crypto_kem/kyber768/clean/params.h b/crypto_kem/kyber768/clean/params.h
index f685061a..bd1dfe14 100644
--- a/crypto_kem/kyber768/clean/params.h
+++ b/crypto_kem/kyber768/clean/params.h
@@ -3,25 +3,29 @@
 
 #define KYBER_K 3
 
+/* Don't change parameters below this line */
+
 #define KYBER_N 256
-#define KYBER_Q 7681
+#define KYBER_Q 3329
 
-#define KYBER_ETA 4
+#define KYBER_ETA 2
 
-#define KYBER_SYMBYTES 32   /* size in bytes of shared key, hashes, and seeds */
+#define KYBER_SYMBYTES 32   /* size in bytes of hashes, and seeds */
+#define KYBER_SSBYTES  32   /* size in bytes of shared key */
 
-#define KYBER_POLYBYTES              416
-#define KYBER_POLYCOMPRESSEDBYTES    96
+#define KYBER_POLYBYTES              384
 #define KYBER_POLYVECBYTES           (KYBER_K * KYBER_POLYBYTES)
-#define KYBER_POLYVECCOMPRESSEDBYTES (KYBER_K * 352)
+
+#define KYBER_POLYCOMPRESSEDBYTES    128
+#define KYBER_POLYVECCOMPRESSEDBYTES (KYBER_K * 320)
 
 #define KYBER_INDCPA_MSGBYTES       KYBER_SYMBYTES
-#define KYBER_INDCPA_PUBLICKEYBYTES (KYBER_POLYVECCOMPRESSEDBYTES + KYBER_SYMBYTES)
+#define KYBER_INDCPA_PUBLICKEYBYTES (KYBER_POLYVECBYTES + KYBER_SYMBYTES)
 #define KYBER_INDCPA_SECRETKEYBYTES (KYBER_POLYVECBYTES)
 #define KYBER_INDCPA_BYTES          (KYBER_POLYVECCOMPRESSEDBYTES + KYBER_POLYCOMPRESSEDBYTES)
 
 #define KYBER_PUBLICKEYBYTES  (KYBER_INDCPA_PUBLICKEYBYTES)
-#define KYBER_SECRETKEYBYTES  (KYBER_INDCPA_SECRETKEYBYTES + KYBER_INDCPA_PUBLICKEYBYTES + 2 * KYBER_SYMBYTES) /* 32 bytes of additional space to save H(pk) */
+#define KYBER_SECRETKEYBYTES  (KYBER_INDCPA_SECRETKEYBYTES +  KYBER_INDCPA_PUBLICKEYBYTES + 2*KYBER_SYMBYTES) /* 32 bytes of additional space to save H(pk) */
 #define KYBER_CIPHERTEXTBYTES  KYBER_INDCPA_BYTES
 
 #endif
diff --git a/crypto_kem/kyber768/clean/poly.c b/crypto_kem/kyber768/clean/poly.c
index 6eb18330..98730297 100644
--- a/crypto_kem/kyber768/clean/poly.c
+++ b/crypto_kem/kyber768/clean/poly.c
@@ -1,210 +1,260 @@
 #include "poly.h"
+
 #include "cbd.h"
-#include "fips202.h"
 #include "ntt.h"
-#include "polyvec.h"
+#include "params.h"
 #include "reduce.h"
-#include <stdio.h>
+#include "symmetric.h"
+
+#include <stdint.h>
 
 /*************************************************
- * Name:        poly_compress
- *
- * Description: Compression and subsequent serialization of a polynomial
- *
- * Arguments:   - unsigned char *r: pointer to output byte array
- *              - const poly *a:    pointer to input polynomial
- **************************************************/
-void PQCLEAN_KYBER768_CLEAN_poly_compress(unsigned char *r, const poly *a) {
-    uint32_t t[8];
-    unsigned int i, j, k = 0;
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_compress
+*
+* Description: Compression and subsequent serialization of a polynomial
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYCOMPRESSEDBYTES bytes)
+*              - const poly *a:    pointer to input polynomial
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_poly_compress(unsigned char *r, poly *a) {
+    uint8_t t[8];
+    int i, j, k = 0;
+
+    PQCLEAN_KYBER768_CLEAN_poly_csubq(a);
 
     for (i = 0; i < KYBER_N; i += 8) {
         for (j = 0; j < 8; j++) {
-            t[j] = (((PQCLEAN_KYBER768_CLEAN_freeze(a->coeffs[i + j]) << 3) + KYBER_Q / 2) / KYBER_Q) & 7;
+            t[j] = ((((uint32_t)a->coeffs[i + j] << 4) + KYBER_Q / 2) / KYBER_Q) & 15;
         }
 
-        r[k]     = (unsigned char)( t[0]       | (t[1] << 3) | (t[2] << 6));
-        r[k + 1] = (unsigned char)((t[2] >> 2) | (t[3] << 1) | (t[4] << 4) | (t[5] << 7));
-        r[k + 2] = (unsigned char)((t[5] >> 1) | (t[6] << 2) | (t[7] << 5));
-        k += 3;
+        r[k]   = t[0] | (t[1] << 4);
+        r[k + 1] = t[2] | (t[3] << 4);
+        r[k + 2] = t[4] | (t[5] << 4);
+        r[k + 3] = t[6] | (t[7] << 4);
+        k += 4;
     }
 }
 
 /*************************************************
- * Name:        poly_decompress
- *
- * Description: De-serialization and subsequent decompression of a polynomial;
- *              approximate inverse of poly_compress
- *
- * Arguments:   - poly *r:                pointer to output polynomial
- *              - const unsigned char *a: pointer to input byte array
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_decompress
+*
+* Description: De-serialization and subsequent decompression of a polynomial;
+*              approximate inverse of PQCLEAN_KYBER768_CLEAN_poly_compress
+*
+* Arguments:   - poly *r:                pointer to output polynomial
+*              - const unsigned char *a: pointer to input byte array (of length KYBER_POLYCOMPRESSEDBYTES bytes)
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_poly_decompress(poly *r, const unsigned char *a) {
-    unsigned int i;
+    int i;
     for (i = 0; i < KYBER_N; i += 8) {
-        r->coeffs[i + 0] =  (((a[0] & 7) * KYBER_Q) + 4) >> 3;
-        r->coeffs[i + 1] = ((((a[0] >> 3) & 7) * KYBER_Q) + 4) >> 3;
-        r->coeffs[i + 2] = ((((a[0] >> 6) | ((a[1] << 2) & 4)) * KYBER_Q) + 4) >> 3;
-        r->coeffs[i + 3] = ((((a[1] >> 1) & 7) * KYBER_Q) + 4) >> 3;
-        r->coeffs[i + 4] = ((((a[1] >> 4) & 7) * KYBER_Q) + 4) >> 3;
-        r->coeffs[i + 5] = ((((a[1] >> 7) | ((a[2] << 1) & 6)) * KYBER_Q) + 4) >> 3;
-        r->coeffs[i + 6] = ((((a[2] >> 2) & 7) * KYBER_Q) + 4) >> 3;
-        r->coeffs[i + 7] = ((((a[2] >> 5)) * KYBER_Q) + 4) >> 3;
-        a += 3;
+        r->coeffs[i + 0] = (((a[0] & 15) * KYBER_Q) + 8) >> 4;
+        r->coeffs[i + 1] = (((a[0] >> 4) * KYBER_Q) + 8) >> 4;
+        r->coeffs[i + 2] = (((a[1] & 15) * KYBER_Q) + 8) >> 4;
+        r->coeffs[i + 3] = (((a[1] >> 4) * KYBER_Q) + 8) >> 4;
+        r->coeffs[i + 4] = (((a[2] & 15) * KYBER_Q) + 8) >> 4;
+        r->coeffs[i + 5] = (((a[2] >> 4) * KYBER_Q) + 8) >> 4;
+        r->coeffs[i + 6] = (((a[3] & 15) * KYBER_Q) + 8) >> 4;
+        r->coeffs[i + 7] = (((a[3] >> 4) * KYBER_Q) + 8) >> 4;
+        a += 4;
     }
 }
 
 /*************************************************
- * Name:        poly_tobytes
- *
- * Description: Serialization of a polynomial
- *
- * Arguments:   - unsigned char *r: pointer to output byte array
- *              - const poly *a:    pointer to input polynomial
- **************************************************/
-void PQCLEAN_KYBER768_CLEAN_poly_tobytes(unsigned char *r, const poly *a) {
-    int i, j;
-    uint16_t t[8];
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_tobytes
+*
+* Description: Serialization of a polynomial
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYBYTES bytes)
+*              - const poly *a:    pointer to input polynomial
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_poly_tobytes(unsigned char *r, poly *a) {
+    int i;
+    uint16_t t0, t1;
 
-    for (i = 0; i < KYBER_N / 8; i++) {
-        for (j = 0; j < 8; j++) {
-            t[j] = PQCLEAN_KYBER768_CLEAN_freeze(a->coeffs[8 * i + j]);
-        }
+    PQCLEAN_KYBER768_CLEAN_poly_csubq(a);
 
-        r[13 * i +  0] = (unsigned char)( t[0] & 0xff);
-        r[13 * i +  1] = (unsigned char)((t[0] >>  8) | ((t[1] & 0x07) << 5));
-        r[13 * i +  2] = (unsigned char)((t[1] >>  3) & 0xff);
-        r[13 * i +  3] = (unsigned char)((t[1] >> 11) | ((t[2] & 0x3f) << 2));
-        r[13 * i +  4] = (unsigned char)((t[2] >>  6) | ((t[3] & 0x01) << 7));
-        r[13 * i +  5] = (unsigned char)((t[3] >>  1) & 0xff);
-        r[13 * i +  6] = (unsigned char)((t[3] >>  9) | ((t[4] & 0x0f) << 4));
-        r[13 * i +  7] = (unsigned char)((t[4] >>  4) & 0xff);
-        r[13 * i +  8] = (unsigned char)((t[4] >> 12) | ((t[5] & 0x7f) << 1));
-        r[13 * i +  9] = (unsigned char)((t[5] >>  7) | ((t[6] & 0x03) << 6));
-        r[13 * i + 10] = (unsigned char)((t[6] >>  2) & 0xff);
-        r[13 * i + 11] = (unsigned char)((t[6] >> 10) | ((t[7] & 0x1f) << 3));
-        r[13 * i + 12] = (unsigned char)((t[7] >>  5));
+    for (i = 0; i < KYBER_N / 2; i++) {
+        t0 = a->coeffs[2 * i];
+        t1 = a->coeffs[2 * i + 1];
+        r[3 * i] = t0 & 0xff;
+        r[3 * i + 1] = (t0 >> 8) | ((t1 & 0xf) << 4);
+        r[3 * i + 2] = (t1 >> 4) & 0xff;
     }
 }
 
 /*************************************************
- * Name:        poly_frombytes
- *
- * Description: De-serialization of a polynomial;
- *              inverse of poly_tobytes
- *
- * Arguments:   - poly *r:                pointer to output polynomial
- *              - const unsigned char *a: pointer to input byte array
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_frombytes
+*
+* Description: De-serialization of a polynomial;
+*              inverse of PQCLEAN_KYBER768_CLEAN_poly_tobytes
+*
+* Arguments:   - poly *r:                pointer to output polynomial
+*              - const unsigned char *a: pointer to input byte array (of KYBER_POLYBYTES bytes)
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_poly_frombytes(poly *r, const unsigned char *a) {
     int i;
-    for (i = 0; i < KYBER_N / 8; i++) {
-        r->coeffs[8 * i + 0]  = a[13 * i + 0]        | (((uint16_t)a[13 * i +  1] & 0x1f) << 8);
-        r->coeffs[8 * i + 1] = (a[13 * i +  1] >> 5) | (((uint16_t)a[13 * i +  2]) << 3) | (((uint16_t)a[13 * i +  3] & 0x03) << 11);
-        r->coeffs[8 * i + 2] = (a[13 * i +  3] >> 2) | (((uint16_t)a[13 * i +  4] & 0x7f) << 6);
-        r->coeffs[8 * i + 3] = (a[13 * i +  4] >> 7) | (((uint16_t)a[13 * i +  5]) << 1) | (((uint16_t)a[13 * i +  6] & 0x0f) << 9);
-        r->coeffs[8 * i + 4] = (a[13 * i +  6] >> 4) | (((uint16_t)a[13 * i +  7]) << 4) | (((uint16_t)a[13 * i +  8] & 0x01) << 12);
-        r->coeffs[8 * i + 5] = (a[13 * i +  8] >> 1) | (((uint16_t)a[13 * i +  9] & 0x3f) << 7);
-        r->coeffs[8 * i + 6] = (a[13 * i +  9] >> 6) | (((uint16_t)a[13 * i + 10]) << 2) | (((uint16_t)a[13 * i + 11] & 0x07) << 10);
-        r->coeffs[8 * i + 7] = (a[13 * i + 11] >> 3) | (((uint16_t)a[13 * i + 12]) << 5);
+
+    for (i = 0; i < KYBER_N / 2; i++) {
+        r->coeffs[2 * i]   = a[3 * i]        | ((uint16_t)a[3 * i + 1] & 0x0f) << 8;
+        r->coeffs[2 * i + 1] = a[3 * i + 1] >> 4 | ((uint16_t)a[3 * i + 2] & 0xff) << 4;
     }
 }
 
 /*************************************************
- * Name:        poly_getnoise
- *
- * Description: Sample a polynomial deterministically from a seed and a nonce,
- *              with output polynomial close to centered binomial distribution
- *              with parameter KYBER_ETA
- *
- * Arguments:   - poly *r:                   pointer to output polynomial
- *              - const unsigned char *seed: pointer to input seed
- *              - unsigned char nonce:       one-byte input nonce
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_getnoise
+*
+* Description: Sample a polynomial deterministically from a seed and a nonce,
+*              with output polynomial close to centered binomial distribution
+*              with parameter KYBER_ETA
+*
+* Arguments:   - poly *r:                   pointer to output polynomial
+*              - const unsigned char *seed: pointer to input seed (pointing to array of length KYBER_SYMBYTES bytes)
+*              - unsigned char nonce:       one-byte input nonce
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_poly_getnoise(poly *r, const unsigned char *seed, unsigned char nonce) {
     unsigned char buf[KYBER_ETA * KYBER_N / 4];
-    unsigned char extseed[KYBER_SYMBYTES + 1];
-    int i;
-
-    for (i = 0; i < KYBER_SYMBYTES; i++) {
-        extseed[i] = seed[i];
-    }
-    extseed[KYBER_SYMBYTES] = nonce;
-
-    shake256(buf, KYBER_ETA * KYBER_N / 4, extseed, KYBER_SYMBYTES + 1);
 
+    prf(buf, KYBER_ETA * KYBER_N / 4, seed, nonce);
     PQCLEAN_KYBER768_CLEAN_cbd(r, buf);
 }
 
 /*************************************************
- * Name:        poly_ntt
- *
- * Description: Computes negacyclic number-theoretic transform (NTT) of
- *              a polynomial in place;
- *              inputs assumed to be in normal order, output in bitreversed
- *order
- *
- * Arguments:   - uint16_t *r: pointer to in/output polynomial
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_ntt
+*
+* Description: Computes negacyclic number-theoretic transform (NTT) of
+*              a polynomial in place;
+*              inputs assumed to be in normal order, output in bitreversed order
+*
+* Arguments:   - uint16_t *r: pointer to in/output polynomial
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_poly_ntt(poly *r) {
     PQCLEAN_KYBER768_CLEAN_ntt(r->coeffs);
+    PQCLEAN_KYBER768_CLEAN_poly_reduce(r);
 }
 
 /*************************************************
- * Name:        poly_invntt
- *
- * Description: Computes inverse of negacyclic number-theoretic transform (NTT)
- *of a polynomial in place; inputs assumed to be in bitreversed order, output in
- *normal order
- *
- * Arguments:   - uint16_t *a: pointer to in/output polynomial
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_invntt
+*
+* Description: Computes inverse of negacyclic number-theoretic transform (NTT) of
+*              a polynomial in place;
+*              inputs assumed to be in bitreversed order, output in normal order
+*
+* Arguments:   - uint16_t *a: pointer to in/output polynomial
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_poly_invntt(poly *r) {
     PQCLEAN_KYBER768_CLEAN_invntt(r->coeffs);
 }
 
 /*************************************************
- * Name:        poly_add
- *
- * Description: Add two polynomials
- *
- * Arguments: - poly *r:       pointer to output polynomial
- *            - const poly *a: pointer to first input polynomial
- *            - const poly *b: pointer to second input polynomial
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_basemul
+*
+* Description: Multiplication of two polynomials in NTT domain
+*
+* Arguments:   - poly *r:       pointer to output polynomial
+*              - const poly *a: pointer to first input polynomial
+*              - const poly *b: pointer to second input polynomial
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_poly_basemul(poly *r, const poly *a, const poly *b) {
+    unsigned int i;
+
+    for (i = 0; i < KYBER_N / 4; ++i) {
+        PQCLEAN_KYBER768_CLEAN_basemul(r->coeffs + 4 * i, a->coeffs + 4 * i, b->coeffs + 4 * i, PQCLEAN_KYBER768_CLEAN_zetas[64 + i]);
+        PQCLEAN_KYBER768_CLEAN_basemul(r->coeffs + 4 * i + 2, a->coeffs + 4 * i + 2, b->coeffs + 4 * i + 2, - PQCLEAN_KYBER768_CLEAN_zetas[64 + i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_frommont
+*
+* Description: Inplace conversion of all coefficients of a polynomial
+*              from Montgomery domain to normal domain
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_poly_frommont(poly *r) {
+    int i;
+    const int16_t f = (1ULL << 32) % KYBER_Q;
+
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = PQCLEAN_KYBER768_CLEAN_montgomery_reduce((int32_t)r->coeffs[i] * f);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_reduce
+*
+* Description: Applies Barrett reduction to all coefficients of a polynomial
+*              for details of the Barrett reduction see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_poly_reduce(poly *r) {
+    int i;
+
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = PQCLEAN_KYBER768_CLEAN_barrett_reduce(r->coeffs[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_csubq
+*
+* Description: Applies conditional subtraction of q to each coefficient of a polynomial
+*              for details of conditional subtraction of q see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_poly_csubq(poly *r) {
+    int i;
+
+    for (i = 0; i < KYBER_N; i++) {
+        r->coeffs[i] = PQCLEAN_KYBER768_CLEAN_csubq(r->coeffs[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_add
+*
+* Description: Add two polynomials
+*
+* Arguments: - poly *r:       pointer to output polynomial
+*            - const poly *a: pointer to first input polynomial
+*            - const poly *b: pointer to second input polynomial
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_poly_add(poly *r, const poly *a, const poly *b) {
     int i;
     for (i = 0; i < KYBER_N; i++) {
-        r->coeffs[i] = PQCLEAN_KYBER768_CLEAN_barrett_reduce(a->coeffs[i] + b->coeffs[i]);
+        r->coeffs[i] = a->coeffs[i] + b->coeffs[i];
     }
 }
 
 /*************************************************
- * Name:        poly_sub
- *
- * Description: Subtract two polynomials
- *
- * Arguments: - poly *r:       pointer to output polynomial
- *            - const poly *a: pointer to first input polynomial
- *            - const poly *b: pointer to second input polynomial
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_sub
+*
+* Description: Subtract two polynomials
+*
+* Arguments: - poly *r:       pointer to output polynomial
+*            - const poly *a: pointer to first input polynomial
+*            - const poly *b: pointer to second input polynomial
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_poly_sub(poly *r, const poly *a, const poly *b) {
     int i;
     for (i = 0; i < KYBER_N; i++) {
-        r->coeffs[i] = PQCLEAN_KYBER768_CLEAN_barrett_reduce(a->coeffs[i] + 3 * KYBER_Q - b->coeffs[i]);
+        r->coeffs[i] = a->coeffs[i] - b->coeffs[i];
     }
 }
 
 /*************************************************
- * Name:        poly_frommsg
- *
- * Description: Convert 32-byte message to polynomial
- *
- * Arguments:   - poly *r:                  pointer to output polynomial
- *              - const unsigned char *msg: pointer to input message
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_frommsg
+*
+* Description: Convert 32-byte message to polynomial
+*
+* Arguments:   - poly *r:                  pointer to output polynomial
+*              - const unsigned char *msg: pointer to input message
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_poly_frommsg(poly *r, const unsigned char msg[KYBER_SYMBYTES]) {
-    uint16_t i, j, mask;
+    int i, j;
+    uint16_t mask;
 
     for (i = 0; i < KYBER_SYMBYTES; i++) {
         for (j = 0; j < 8; j++) {
@@ -215,21 +265,23 @@ void PQCLEAN_KYBER768_CLEAN_poly_frommsg(poly *r, const unsigned char msg[KYBER_
 }
 
 /*************************************************
- * Name:        poly_tomsg
- *
- * Description: Convert polynomial to 32-byte message
- *
- * Arguments:   - unsigned char *msg: pointer to output message
- *              - const poly *a:      pointer to input polynomial
- **************************************************/
-void PQCLEAN_KYBER768_CLEAN_poly_tomsg(unsigned char msg[KYBER_SYMBYTES], const poly *a) {
+* Name:        PQCLEAN_KYBER768_CLEAN_poly_tomsg
+*
+* Description: Convert polynomial to 32-byte message
+*
+* Arguments:   - unsigned char *msg: pointer to output message
+*              - const poly *a:      pointer to input polynomial
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_poly_tomsg(unsigned char msg[KYBER_SYMBYTES], poly *a) {
     uint16_t t;
     int i, j;
 
+    PQCLEAN_KYBER768_CLEAN_poly_csubq(a);
+
     for (i = 0; i < KYBER_SYMBYTES; i++) {
         msg[i] = 0;
         for (j = 0; j < 8; j++) {
-            t = (((PQCLEAN_KYBER768_CLEAN_freeze(a->coeffs[8 * i + j]) << 1) + KYBER_Q / 2) / KYBER_Q) & 1;
+            t = (((a->coeffs[8 * i + j] << 1) + KYBER_Q / 2) / KYBER_Q) & 1;
             msg[i] |= t << j;
         }
     }
diff --git a/crypto_kem/kyber768/clean/poly.h b/crypto_kem/kyber768/clean/poly.h
index 9b9f2968..4529c2dc 100644
--- a/crypto_kem/kyber768/clean/poly.h
+++ b/crypto_kem/kyber768/clean/poly.h
@@ -2,6 +2,7 @@
 #define POLY_H
 
 #include "params.h"
+
 #include <stdint.h>
 
 /*
@@ -9,22 +10,27 @@
  * coeffs[0] + X*coeffs[1] + X^2*xoeffs[2] + ... + X^{n-1}*coeffs[n-1]
  */
 typedef struct {
-    uint16_t coeffs[KYBER_N];
+    int16_t coeffs[KYBER_N];
 } poly;
 
-void PQCLEAN_KYBER768_CLEAN_poly_compress(unsigned char *r, const poly *a);
+void PQCLEAN_KYBER768_CLEAN_poly_compress(unsigned char *r, poly *a);
 void PQCLEAN_KYBER768_CLEAN_poly_decompress(poly *r, const unsigned char *a);
 
-void PQCLEAN_KYBER768_CLEAN_poly_tobytes(unsigned char *r, const poly *a);
+void PQCLEAN_KYBER768_CLEAN_poly_tobytes(unsigned char *r, poly *a);
 void PQCLEAN_KYBER768_CLEAN_poly_frombytes(poly *r, const unsigned char *a);
 
 void PQCLEAN_KYBER768_CLEAN_poly_frommsg(poly *r, const unsigned char msg[KYBER_SYMBYTES]);
-void PQCLEAN_KYBER768_CLEAN_poly_tomsg(unsigned char msg[KYBER_SYMBYTES], const poly *a);
+void PQCLEAN_KYBER768_CLEAN_poly_tomsg(unsigned char msg[KYBER_SYMBYTES], poly *a);
 
 void PQCLEAN_KYBER768_CLEAN_poly_getnoise(poly *r, const unsigned char *seed, unsigned char nonce);
 
 void PQCLEAN_KYBER768_CLEAN_poly_ntt(poly *r);
 void PQCLEAN_KYBER768_CLEAN_poly_invntt(poly *r);
+void PQCLEAN_KYBER768_CLEAN_poly_basemul(poly *r, const poly *a, const poly *b);
+void PQCLEAN_KYBER768_CLEAN_poly_frommont(poly *r);
+
+void PQCLEAN_KYBER768_CLEAN_poly_reduce(poly *r);
+void PQCLEAN_KYBER768_CLEAN_poly_csubq(poly *r);
 
 void PQCLEAN_KYBER768_CLEAN_poly_add(poly *r, const poly *a, const poly *b);
 void PQCLEAN_KYBER768_CLEAN_poly_sub(poly *r, const poly *a, const poly *b);
diff --git a/crypto_kem/kyber768/clean/polyvec.c b/crypto_kem/kyber768/clean/polyvec.c
index 147a3854..91c35379 100644
--- a/crypto_kem/kyber768/clean/polyvec.c
+++ b/crypto_kem/kyber768/clean/polyvec.c
@@ -1,77 +1,70 @@
 #include "polyvec.h"
-#include "cbd.h"
-#include "fips202.h"
-#include "reduce.h"
-#include <stdio.h>
+
+#include "poly.h"
+
+#include <stdint.h>
 
 /*************************************************
- * Name:        polyvec_compress
- *
- * Description: Compress and serialize vector of polynomials
- *
- * Arguments:   - unsigned char *r: pointer to output byte array
- *              - const polyvec *a: pointer to input vector of polynomials
- **************************************************/
-void PQCLEAN_KYBER768_CLEAN_polyvec_compress(unsigned char *r, const polyvec *a) {
+* Name:        PQCLEAN_KYBER768_CLEAN_polyvec_compress
+*
+* Description: Compress and serialize vector of polynomials
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYVECCOMPRESSEDBYTES)
+*              - const polyvec *a: pointer to input vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_polyvec_compress(unsigned char *r, polyvec *a) {
     int i, j, k;
-    uint16_t t[8];
+
+    PQCLEAN_KYBER768_CLEAN_polyvec_csubq(a);
+
+    uint16_t t[4];
     for (i = 0; i < KYBER_K; i++) {
-        for (j = 0; j < KYBER_N / 8; j++) {
-            for (k = 0; k < 8; k++) {
-                t[k] = ((((uint32_t)PQCLEAN_KYBER768_CLEAN_freeze(a->vec[i].coeffs[8 * j + k]) << 11) + KYBER_Q / 2) / KYBER_Q) & 0x7ff;
+        for (j = 0; j < KYBER_N / 4; j++) {
+            for (k = 0; k < 4; k++) {
+                t[k] = ((((uint32_t)a->vec[i].coeffs[4 * j + k] << 10) + KYBER_Q / 2) / KYBER_Q) & 0x3ff;
             }
 
-            r[11 * j +  0] = (unsigned char)( t[0] & 0xff);
-            r[11 * j +  1] = (unsigned char)((t[0] >>  8) | ((t[1] & 0x1f) << 3));
-            r[11 * j +  2] = (unsigned char)((t[1] >>  5) | ((t[2] & 0x03) << 6));
-            r[11 * j +  3] = (unsigned char)((t[2] >>  2) & 0xff);
-            r[11 * j +  4] = (unsigned char)((t[2] >> 10) | ((t[3] & 0x7f) << 1));
-            r[11 * j +  5] = (unsigned char)((t[3] >>  7) | ((t[4] & 0x0f) << 4));
-            r[11 * j +  6] = (unsigned char)((t[4] >>  4) | ((t[5] & 0x01) << 7));
-            r[11 * j +  7] = (unsigned char)((t[5] >>  1) & 0xff);
-            r[11 * j +  8] = (unsigned char)((t[5] >>  9) | ((t[6] & 0x3f) << 2));
-            r[11 * j +  9] = (unsigned char)((t[6] >>  6) | ((t[7] & 0x07) << 5));
-            r[11 * j + 10] = (unsigned char)((t[7] >>  3));
+            r[5 * j + 0] =  t[0] & 0xff;
+            r[5 * j + 1] = (t[0] >>  8) | ((t[1] & 0x3f) << 2);
+            r[5 * j + 2] = ((t[1] >>  6) | ((t[2] & 0x0f) << 4)) & 0xff;
+            r[5 * j + 3] = ((t[2] >>  4) | ((t[3] & 0x03) << 6)) & 0xff;
+            r[5 * j + 4] = (t[3] >>  2) & 0xff;
         }
-        r += 352;
+        r += 320;
     }
 }
 
 /*************************************************
- * Name:        polyvec_decompress
- *
- * Description: De-serialize and decompress vector of polynomials;
- *              approximate inverse of polyvec_compress
- *
- * Arguments:   - polyvec *r:       pointer to output vector of polynomials
- *              - unsigned char *a: pointer to input byte array
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_polyvec_decompress
+*
+* Description: De-serialize and decompress vector of polynomials;
+*              approximate inverse of PQCLEAN_KYBER768_CLEAN_polyvec_compress
+*
+* Arguments:   - polyvec *r:       pointer to output vector of polynomials
+*              - unsigned char *a: pointer to input byte array (of length KYBER_POLYVECCOMPRESSEDBYTES)
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_polyvec_decompress(polyvec *r, const unsigned char *a) {
     int i, j;
     for (i = 0; i < KYBER_K; i++) {
-        for (j = 0; j < KYBER_N / 8; j++) {
-            r->vec[i].coeffs[8 * j + 0] =  (((a[11 * j + 0]       | (((uint32_t)a[11 * j +  1] & 0x07) << 8)) * KYBER_Q) + 1024) >> 11;
-            r->vec[i].coeffs[8 * j + 1] = ((((a[11 * j + 1] >> 3) | (((uint32_t)a[11 * j +  2] & 0x3f) << 5)) * KYBER_Q) + 1024) >> 11;
-            r->vec[i].coeffs[8 * j + 2] = ((((a[11 * j + 2] >> 6) | (((uint32_t)a[11 * j +  3] & 0xff) << 2) | (((uint32_t)a[11 * j + 4] & 0x01) << 10)) * KYBER_Q) + 1024) >> 11;
-            r->vec[i].coeffs[8 * j + 3] = ((((a[11 * j + 4] >> 1) | (((uint32_t)a[11 * j +  5] & 0x0f) << 7)) * KYBER_Q) + 1024) >> 11;
-            r->vec[i].coeffs[8 * j + 4] = ((((a[11 * j + 5] >> 4) | (((uint32_t)a[11 * j +  6] & 0x7f) << 4)) * KYBER_Q) + 1024) >> 11;
-            r->vec[i].coeffs[8 * j + 5] = ((((a[11 * j + 6] >> 7) | (((uint32_t)a[11 * j +  7] & 0xff) << 1) | (((uint32_t)a[11 * j + 8] & 0x03) <<  9)) * KYBER_Q) + 1024) >> 11;
-            r->vec[i].coeffs[8 * j + 6] = ((((a[11 * j + 8] >> 2) | (((uint32_t)a[11 * j +  9] & 0x1f) << 6)) * KYBER_Q) + 1024) >> 11;
-            r->vec[i].coeffs[8 * j + 7] = ((((a[11 * j + 9] >> 5) | (((uint32_t)a[11 * j + 10] & 0xff) << 3)) * KYBER_Q) + 1024) >> 11;
+        for (j = 0; j < KYBER_N / 4; j++) {
+            r->vec[i].coeffs[4 * j + 0] =  (((a[5 * j + 0]       | (((uint32_t)a[5 * j + 1] & 0x03) << 8)) * KYBER_Q) + 512) >> 10;
+            r->vec[i].coeffs[4 * j + 1] = ((((a[5 * j + 1] >> 2) | (((uint32_t)a[5 * j + 2] & 0x0f) << 6)) * KYBER_Q) + 512) >> 10;
+            r->vec[i].coeffs[4 * j + 2] = ((((a[5 * j + 2] >> 4) | (((uint32_t)a[5 * j + 3] & 0x3f) << 4)) * KYBER_Q) + 512) >> 10;
+            r->vec[i].coeffs[4 * j + 3] = ((((a[5 * j + 3] >> 6) | (((uint32_t)a[5 * j + 4] & 0xff) << 2)) * KYBER_Q) + 512) >> 10;
         }
-        a += 352;
+        a += 320;
     }
 }
 
 /*************************************************
- * Name:        polyvec_tobytes
- *
- * Description: Serialize vector of polynomials
- *
- * Arguments:   - unsigned char *r: pointer to output byte array
- *              - const polyvec *a: pointer to input vector of polynomials
- **************************************************/
-void PQCLEAN_KYBER768_CLEAN_polyvec_tobytes(unsigned char *r, const polyvec *a) {
+* Name:        PQCLEAN_KYBER768_CLEAN_polyvec_tobytes
+*
+* Description: Serialize vector of polynomials
+*
+* Arguments:   - unsigned char *r: pointer to output byte array (needs space for KYBER_POLYVECBYTES)
+*              - const polyvec *a: pointer to input vector of polynomials
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_polyvec_tobytes(unsigned char *r, polyvec *a) {
     int i;
     for (i = 0; i < KYBER_K; i++) {
         PQCLEAN_KYBER768_CLEAN_poly_tobytes(r + i * KYBER_POLYBYTES, &a->vec[i]);
@@ -79,14 +72,14 @@ void PQCLEAN_KYBER768_CLEAN_polyvec_tobytes(unsigned char *r, const polyvec *a)
 }
 
 /*************************************************
- * Name:        polyvec_frombytes
- *
- * Description: De-serialize vector of polynomials;
- *              inverse of polyvec_tobytes
- *
- * Arguments:   - unsigned char *r: pointer to output byte array
- *              - const polyvec *a: pointer to input vector of polynomials
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_polyvec_frombytes
+*
+* Description: De-serialize vector of polynomials;
+*              inverse of PQCLEAN_KYBER768_CLEAN_polyvec_tobytes
+*
+* Arguments:   - unsigned char *r: pointer to output byte array
+*              - const polyvec *a: pointer to input vector of polynomials (of length KYBER_POLYVECBYTES)
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_polyvec_frombytes(polyvec *r, const unsigned char *a) {
     int i;
     for (i = 0; i < KYBER_K; i++) {
@@ -95,12 +88,12 @@ void PQCLEAN_KYBER768_CLEAN_polyvec_frombytes(polyvec *r, const unsigned char *a
 }
 
 /*************************************************
- * Name:        polyvec_ntt
- *
- * Description: Apply forward NTT to all elements of a vector of polynomials
- *
- * Arguments:   - polyvec *r: pointer to in/output vector of polynomials
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_polyvec_ntt
+*
+* Description: Apply forward NTT to all elements of a vector of polynomials
+*
+* Arguments:   - polyvec *r: pointer to in/output vector of polynomials
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_polyvec_ntt(polyvec *r) {
     int i;
     for (i = 0; i < KYBER_K; i++) {
@@ -109,12 +102,12 @@ void PQCLEAN_KYBER768_CLEAN_polyvec_ntt(polyvec *r) {
 }
 
 /*************************************************
- * Name:        polyvec_invntt
- *
- * Description: Apply inverse NTT to all elements of a vector of polynomials
- *
- * Arguments:   - polyvec *r: pointer to in/output vector of polynomials
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_polyvec_invntt
+*
+* Description: Apply inverse NTT to all elements of a vector of polynomials
+*
+* Arguments:   - polyvec *r: pointer to in/output vector of polynomials
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_polyvec_invntt(polyvec *r) {
     int i;
     for (i = 0; i < KYBER_K; i++) {
@@ -123,37 +116,68 @@ void PQCLEAN_KYBER768_CLEAN_polyvec_invntt(polyvec *r) {
 }
 
 /*************************************************
- * Name:        polyvec_pointwise_acc
- *
- * Description: Pointwise multiply elements of a and b and accumulate into r
- *
- * Arguments: - poly *r:          pointer to output polynomial
- *            - const polyvec *a: pointer to first input vector of polynomials
- *            - const polyvec *b: pointer to second input vector of polynomials
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_polyvec_pointwise_acc
+*
+* Description: Pointwise multiply elements of a and b and accumulate into r
+*
+* Arguments: - poly *r:          pointer to output polynomial
+*            - const polyvec *a: pointer to first input vector of polynomials
+*            - const polyvec *b: pointer to second input vector of polynomials
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_polyvec_pointwise_acc(poly *r, const polyvec *a, const polyvec *b) {
-    int i, j;
-    uint16_t t;
-    for (j = 0; j < KYBER_N; j++) {
-        t = PQCLEAN_KYBER768_CLEAN_montgomery_reduce(4613 * (uint32_t)b->vec[0].coeffs[j]);  // 4613 = 2^{2*18} % q
-        r->coeffs[j] = PQCLEAN_KYBER768_CLEAN_montgomery_reduce(a->vec[0].coeffs[j] * t);
-        for (i = 1; i < KYBER_K; i++) {
-            t = PQCLEAN_KYBER768_CLEAN_montgomery_reduce(4613 * (uint32_t)b->vec[i].coeffs[j]);
-            r->coeffs[j] += PQCLEAN_KYBER768_CLEAN_montgomery_reduce(a->vec[i].coeffs[j] * t);
-        }
-        r->coeffs[j] = PQCLEAN_KYBER768_CLEAN_barrett_reduce(r->coeffs[j]);
+    int i;
+    poly t;
+
+    PQCLEAN_KYBER768_CLEAN_poly_basemul(r, &a->vec[0], &b->vec[0]);
+    for (i = 1; i < KYBER_K; i++) {
+        PQCLEAN_KYBER768_CLEAN_poly_basemul(&t, &a->vec[i], &b->vec[i]);
+        PQCLEAN_KYBER768_CLEAN_poly_add(r, r, &t);
+    }
+
+    PQCLEAN_KYBER768_CLEAN_poly_reduce(r);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER768_CLEAN_polyvec_reduce
+*
+* Description: Applies Barrett reduction to each coefficient
+*              of each element of a vector of polynomials
+*              for details of the Barrett reduction see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_polyvec_reduce(polyvec *r) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER768_CLEAN_poly_reduce(&r->vec[i]);
     }
 }
 
 /*************************************************
- * Name:        polyvec_add
- *
- * Description: Add vectors of polynomials
- *
- * Arguments: - polyvec *r:       pointer to output vector of polynomials
- *            - const polyvec *a: pointer to first input vector of polynomials
- *            - const polyvec *b: pointer to second input vector of polynomials
- **************************************************/
+* Name:        PQCLEAN_KYBER768_CLEAN_polyvec_csubq
+*
+* Description: Applies conditional subtraction of q to each coefficient
+*              of each element of a vector of polynomials
+*              for details of conditional subtraction of q see comments in reduce.c
+*
+* Arguments:   - poly *r:       pointer to input/output polynomial
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_polyvec_csubq(polyvec *r) {
+    int i;
+    for (i = 0; i < KYBER_K; i++) {
+        PQCLEAN_KYBER768_CLEAN_poly_csubq(&r->vec[i]);
+    }
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER768_CLEAN_polyvec_add
+*
+* Description: Add vectors of polynomials
+*
+* Arguments: - polyvec *r:       pointer to output vector of polynomials
+*            - const polyvec *a: pointer to first input vector of polynomials
+*            - const polyvec *b: pointer to second input vector of polynomials
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_polyvec_add(polyvec *r, const polyvec *a, const polyvec *b) {
     int i;
     for (i = 0; i < KYBER_K; i++) {
diff --git a/crypto_kem/kyber768/clean/polyvec.h b/crypto_kem/kyber768/clean/polyvec.h
index c6fb61ad..41134e26 100644
--- a/crypto_kem/kyber768/clean/polyvec.h
+++ b/crypto_kem/kyber768/clean/polyvec.h
@@ -8,10 +8,10 @@ typedef struct {
     poly vec[KYBER_K];
 } polyvec;
 
-void PQCLEAN_KYBER768_CLEAN_polyvec_compress(unsigned char *r, const polyvec *a);
+void PQCLEAN_KYBER768_CLEAN_polyvec_compress(unsigned char *r, polyvec *a);
 void PQCLEAN_KYBER768_CLEAN_polyvec_decompress(polyvec *r, const unsigned char *a);
 
-void PQCLEAN_KYBER768_CLEAN_polyvec_tobytes(unsigned char *r, const polyvec *a);
+void PQCLEAN_KYBER768_CLEAN_polyvec_tobytes(unsigned char *r, polyvec *a);
 void PQCLEAN_KYBER768_CLEAN_polyvec_frombytes(polyvec *r, const unsigned char *a);
 
 void PQCLEAN_KYBER768_CLEAN_polyvec_ntt(polyvec *r);
@@ -19,6 +19,9 @@ void PQCLEAN_KYBER768_CLEAN_polyvec_invntt(polyvec *r);
 
 void PQCLEAN_KYBER768_CLEAN_polyvec_pointwise_acc(poly *r, const polyvec *a, const polyvec *b);
 
+void PQCLEAN_KYBER768_CLEAN_polyvec_reduce(polyvec *r);
+void PQCLEAN_KYBER768_CLEAN_polyvec_csubq(polyvec *r);
+
 void PQCLEAN_KYBER768_CLEAN_polyvec_add(polyvec *r, const polyvec *a, const polyvec *b);
 
 #endif
diff --git a/crypto_kem/kyber768/clean/precomp.c b/crypto_kem/kyber768/clean/precomp.c
deleted file mode 100644
index 59f158b8..00000000
--- a/crypto_kem/kyber768/clean/precomp.c
+++ /dev/null
@@ -1,100 +0,0 @@
-#include "inttypes.h"
-#include "ntt.h"
-#include "params.h"
-
-/* Precomputed constants for the forward NTT and inverse NTT.
- * Computed using Pari/GP as follows:
- *
-brv=[0,128,64,192,32,160,96,224,16,144,80,208,48,176,112,240, \
-     8,136,72,200,40,168,104,232,24,152,88,216,56,184,120,248, \
-     4,132,68,196,36,164,100,228,20,148,84,212,52,180,116,244, \
-     12,140,76,204,44,172,108,236,28,156,92,220,60,188,124,252, \
-     2,130,66,194,34,162,98,226,18,146,82,210,50,178,114,242, \
-     10,138,74,202,42,170,106,234,26,154,90,218,58,186,122,250, \
-     6,134,70,198,38,166,102,230,22,150,86,214,54,182,118,246, \
-     14,142,78,206,46,174,110,238,30,158,94,222,62,190,126,254, \
-     1,129,65,193,33,161,97,225,17,145,81,209,49,177,113,241, \
-     9,137,73,201,41,169,105,233,25,153,89,217,57,185,121,249, \
-     5,133,69,197,37,165,101,229,21,149,85,213,53,181,117,245, \
-     13,141,77,205,45,173,109,237,29,157,93,221,61,189,125,253, \
-     3,131,67,195,35,163,99,227,19,147,83,211,51,179,115,243, \
-     11,139,75,203,43,171,107,235,27,155,91,219,59,187,123,251, \
-     7,135,71,199,39,167,103,231,23,151,87,215,55,183,119,247, \
-     15,143,79,207,47,175,111,239,31,159,95,223,63,191,127,255];
-
-q = 7681;
-n = 256;
-mont = Mod(2^18,q);
-
-g=0; for(i=2,q-1,if(znorder(Mod(i,q)) == 2*n, g=Mod(i,q); break))
-
-zetas = lift(vector(n, i, g^(brv[i])*mont))
-omegas_inv_bitrev_montgomery = lift(vector(n/2, i,
-(g^2)^(-brv[2*(i-1)+1])*mont)) psis_inv_montgomery = lift(vector(n, i,
-g^(-(i-1))/n*mont))
-
-*/
-
-const uint16_t PQCLEAN_KYBER768_CLEAN_zetas[KYBER_N] = {
-    990,  7427, 2634, 6819, 578,  3281, 2143, 1095, 484,  6362, 3336, 5382,
-    6086, 3823, 877,  5656, 3583, 7010, 6414, 263,  1285, 291,  7143, 7338,
-    1581, 5134, 5184, 5932, 4042, 5775, 2468, 3,    606,  729,  5383, 962,
-    3240, 7548, 5129, 7653, 5929, 4965, 2461, 641,  1584, 2666, 1142, 157,
-    7407, 5222, 5602, 5142, 6140, 5485, 4931, 1559, 2085, 5284, 2056, 3538,
-    7269, 3535, 7190, 1957, 3465, 6792, 1538, 4664, 2023, 7643, 3660, 7673,
-    1694, 6905, 3995, 3475, 5939, 1859, 6910, 4434, 1019, 1492, 7087, 4761,
-    657,  4859, 5798, 2640, 1693, 2607, 2782, 5400, 6466, 1010, 957,  3851,
-    2121, 6392, 7319, 3367, 3659, 3375, 6430, 7583, 1549, 5856, 4773, 6084,
-    5544, 1650, 3997, 4390, 6722, 2915, 4245, 2635, 6128, 7676, 5737, 1616,
-    3457, 3132, 7196, 4702, 6239, 851,  2122, 3009, 7613, 7295, 2007, 323,
-    5112, 3716, 2289, 6442, 6965, 2713, 7126, 3401, 963,  6596, 607,  5027,
-    7078, 4484, 5937, 944,  2860, 2680, 5049, 1777, 5850, 3387, 6487, 6777,
-    4812, 4724, 7077, 186,  6848, 6793, 3463, 5877, 1174, 7116, 3077, 5945,
-    6591, 590,  6643, 1337, 6036, 3991, 1675, 2053, 6055, 1162, 1679, 3883,
-    4311, 2106, 6163, 4486, 6374, 5006, 4576, 4288, 5180, 4102, 282,  6119,
-    7443, 6330, 3184, 4971, 2530, 5325, 4171, 7185, 5175, 5655, 1898, 382,
-    7211, 43,   5965, 6073, 1730, 332,  1577, 3304, 2329, 1699, 6150, 2379,
-    5113, 333,  3502, 4517, 1480, 1172, 5567, 651,  925,  4573, 599,  1367,
-    4109, 1863, 6929, 1605, 3866, 2065, 4048, 839,  5764, 2447, 2022, 3345,
-    1990, 4067, 2036, 2069, 3567, 7371, 2368, 339,  6947, 2159, 654,  7327,
-    2768, 6676, 987,  2214
-};
-
-const uint16_t PQCLEAN_KYBER768_CLEAN_omegas_inv_bitrev_montgomery[KYBER_N / 2] = {
-    990,  254,  862,  5047, 6586, 5538, 4400, 7103, 2025, 6804, 3858, 1595,
-    2299, 4345, 1319, 7197, 7678, 5213, 1906, 3639, 1749, 2497, 2547, 6100,
-    343,  538,  7390, 6396, 7418, 1267, 671,  4098, 5724, 491,  4146, 412,
-    4143, 5625, 2397, 5596, 6122, 2750, 2196, 1541, 2539, 2079, 2459, 274,
-    7524, 6539, 5015, 6097, 7040, 5220, 2716, 1752, 28,   2552, 133,  4441,
-    6719, 2298, 6952, 7075, 4672, 5559, 6830, 1442, 2979, 485,  4549, 4224,
-    6065, 1944, 5,    1553, 5046, 3436, 4766, 959,  3291, 3684, 6031, 2137,
-    1597, 2908, 1825, 6132, 98,   1251, 4306, 4022, 4314, 362,  1289, 5560,
-    3830, 6724, 6671, 1215, 2281, 4899, 5074, 5988, 5041, 1883, 2822, 7024,
-    2920, 594,  6189, 6662, 3247, 771,  5822, 1742, 4206, 3686, 776,  5987,
-    8,    4021, 38,   5658, 3017, 6143, 889,  4216
-};
-
-const uint16_t PQCLEAN_KYBER768_CLEAN_psis_inv_montgomery[KYBER_N] = {
-    1024, 4972, 5779, 6907, 4943, 4168, 315,  5580, 90,   497,  1123, 142,
-    4710, 5527, 2443, 4871, 698,  2489, 2394, 4003, 684,  2241, 2390, 7224,
-    5072, 2064, 4741, 1687, 6841, 482,  7441, 1235, 2126, 4742, 2802, 5744,
-    6287, 4933, 699,  3604, 1297, 2127, 5857, 1705, 3868, 3779, 4397, 2177,
-    159,  622,  2240, 1275, 640,  6948, 4572, 5277, 209,  2605, 1157, 7328,
-    5817, 3191, 1662, 2009, 4864, 574,  2487, 164,  6197, 4436, 7257, 3462,
-    4268, 4281, 3414, 4515, 3170, 1290, 2003, 5855, 7156, 6062, 7531, 1732,
-    3249, 4884, 7512, 3590, 1049, 2123, 1397, 6093, 3691, 6130, 6541, 3946,
-    6258, 3322, 1788, 4241, 4900, 2309, 1400, 1757, 400,  502,  6698, 2338,
-    3011, 668,  7444, 4580, 6516, 6795, 2959, 4136, 3040, 2279, 6355, 3943,
-    2913, 6613, 7416, 4084, 6508, 5556, 4054, 3782, 61,   6567, 2212, 779,
-    632,  5709, 5667, 4923, 4911, 6893, 4695, 4164, 3536, 2287, 7594, 2848,
-    3267, 1911, 3128, 546,  1991, 156,  4958, 5531, 6903, 483,  875,  138,
-    250,  2234, 2266, 7222, 2842, 4258, 812,  6703, 232,  5207, 6650, 2585,
-    1900, 6225, 4932, 7265, 4701, 3173, 4635, 6393, 227,  7313, 4454, 4284,
-    6759, 1224, 5223, 1447, 395,  2608, 4502, 4037, 189,  3348, 54,   6443,
-    2210, 6230, 2826, 1780, 3002, 5995, 1955, 6102, 6045, 3938, 5019, 4417,
-    1434, 1262, 1507, 5847, 5917, 7157, 7177, 6434, 7537, 741,  4348, 1309,
-    145,  374,  2236, 4496, 5028, 6771, 6923, 7421, 1978, 1023, 3857, 6876,
-    1102, 7451, 4704, 6518, 1344, 765,  384,  5705, 1207, 1630, 4734, 1563,
-    6839, 5933, 1954, 4987, 7142, 5814, 7527, 4953, 7637, 4707, 2182, 5734,
-    2818, 541,  4097, 5641
-};
diff --git a/crypto_kem/kyber768/clean/reduce.c b/crypto_kem/kyber768/clean/reduce.c
index 8b023929..572f7e34 100644
--- a/crypto_kem/kyber768/clean/reduce.c
+++ b/crypto_kem/kyber768/clean/reduce.c
@@ -1,70 +1,62 @@
 #include "reduce.h"
+
 #include "params.h"
 
-static const uint32_t qinv = 7679; // -inverse_mod(q,2^18)
-static const uint32_t rlog = 18;
+#include <stdint.h>
 
 /*************************************************
- * Name:        montgomery_reduce
- *
- * Description: Montgomery reduction; given a 32-bit integer a, computes
- *              16-bit integer congruent to a * R^-1 mod q,
- *              where R=2^18 (see value of rlog)
- *
- * Arguments:   - uint32_t a: input unsigned integer to be reduced; has to be in
- *{0,...,2281446912}
- *
- * Returns:     unsigned integer in {0,...,2^13-1} congruent to a * R^-1 modulo
- *q.
- **************************************************/
-uint16_t PQCLEAN_KYBER768_CLEAN_montgomery_reduce(uint32_t a) {
-    uint32_t u;
+* Name:        PQCLEAN_KYBER768_CLEAN_montgomery_reduce
+*
+* Description: Montgomery reduction; given a 32-bit integer a, computes
+*              16-bit integer congruent to a * R^-1 mod q,
+*              where R=2^16
+*
+* Arguments:   - int32_t a: input integer to be reduced; has to be in {-q2^15,...,q2^15-1}
+*
+* Returns:     integer in {-q+1,...,q-1} congruent to a * R^-1 modulo q.
+**************************************************/
+int16_t PQCLEAN_KYBER768_CLEAN_montgomery_reduce(int32_t a) {
+    int32_t t;
+    int16_t u;
 
-    u = (a * qinv);
-    u &= ((1 << rlog) - 1);
-    u *= KYBER_Q;
-    a = a + u;
-    return (uint16_t)(a >> rlog);
+    u = (int16_t)(a * QINV);
+    t = (int32_t)u * KYBER_Q;
+    t = a - t;
+    t >>= 16;
+    return (int16_t)t;
 }
 
 /*************************************************
- * Name:        barrett_reduce
- *
- * Description: Barrett reduction; given a 16-bit integer a, computes
- *              16-bit integer congruent to a mod q in {0,...,11768}
- *
- * Arguments:   - uint16_t a: input unsigned integer to be reduced
- *
- * Returns:     unsigned integer in {0,...,11768} congruent to a modulo q.
- **************************************************/
-uint16_t PQCLEAN_KYBER768_CLEAN_barrett_reduce(uint16_t a) {
-    uint16_t u;
+* Name:        PQCLEAN_KYBER768_CLEAN_barrett_reduce
+*
+* Description: Barrett reduction; given a 16-bit integer a, computes
+*              16-bit integer congruent to a mod q in {0,...,q}
+*
+* Arguments:   - int16_t a: input integer to be reduced
+*
+* Returns:     integer in {0,...,q} congruent to a modulo q.
+**************************************************/
+int16_t PQCLEAN_KYBER768_CLEAN_barrett_reduce(int16_t a) {
+    int32_t t;
+    const int32_t v = (1U << 26) / KYBER_Q + 1;
 
-    u = a >> 13; //((uint32_t) a * sinv) >> 16;
-    u *= KYBER_Q;
-    a -= u;
+    t = v * a;
+    t >>= 26;
+    t *= KYBER_Q;
+    return (int16_t)(a - t);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER768_CLEAN_csubq
+*
+* Description: Conditionallly subtract q
+*
+* Arguments:   - int16_t x: input integer
+*
+* Returns:     a - q if a >= q, else a
+**************************************************/
+int16_t PQCLEAN_KYBER768_CLEAN_csubq(int16_t a) {
+    a -= KYBER_Q;
+    a += (a >> 15) & KYBER_Q;
     return a;
 }
-
-/*************************************************
- * Name:        freeze
- *
- * Description: Full reduction; given a 16-bit integer a, computes
- *              unsigned integer a mod q.
- *
- * Arguments:   - uint16_t x: input unsigned integer to be reduced
- *
- * Returns:     unsigned integer in {0,...,q-1} congruent to a modulo q.
- **************************************************/
-uint16_t PQCLEAN_KYBER768_CLEAN_freeze(uint16_t x) {
-    uint16_t m, r;
-    int16_t c;
-    r = PQCLEAN_KYBER768_CLEAN_barrett_reduce(x);
-
-    m = r - KYBER_Q;
-    c = m;
-    c >>= 15;
-    r = m ^ ((r ^ m) & c);
-
-    return r;
-}
diff --git a/crypto_kem/kyber768/clean/reduce.h b/crypto_kem/kyber768/clean/reduce.h
index e7c4f0be..7a679969 100644
--- a/crypto_kem/kyber768/clean/reduce.h
+++ b/crypto_kem/kyber768/clean/reduce.h
@@ -3,10 +3,13 @@
 
 #include <stdint.h>
 
-uint16_t PQCLEAN_KYBER768_CLEAN_freeze(uint16_t x);
+#define MONT 2285 // 2^16 % Q
+#define QINV 62209 // q^(-1) mod 2^16
 
-uint16_t PQCLEAN_KYBER768_CLEAN_montgomery_reduce(uint32_t a);
+int16_t PQCLEAN_KYBER768_CLEAN_montgomery_reduce(int32_t a);
 
-uint16_t PQCLEAN_KYBER768_CLEAN_barrett_reduce(uint16_t a);
+int16_t PQCLEAN_KYBER768_CLEAN_barrett_reduce(int16_t a);
+
+int16_t PQCLEAN_KYBER768_CLEAN_csubq(int16_t a);
 
 #endif
diff --git a/crypto_kem/kyber768/clean/symmetric-fips202.c b/crypto_kem/kyber768/clean/symmetric-fips202.c
new file mode 100644
index 00000000..3d4e858c
--- /dev/null
+++ b/crypto_kem/kyber768/clean/symmetric-fips202.c
@@ -0,0 +1,64 @@
+#include "fips202.h"
+#include "symmetric.h"
+
+#include <stdlib.h>
+
+/*************************************************
+* Name:        PQCLEAN_KYBER768_CLEAN_kyber_shake128_absorb
+*
+* Description: Absorb step of the SHAKE128 specialized for the Kyber context.
+*
+* Arguments:   - uint64_t *s:                     pointer to (uninitialized) output Keccak state
+*              - const unsigned char *input:      pointer to KYBER_SYMBYTES input to be absorbed into s
+*              - unsigned char i                  additional byte of input
+*              - unsigned char j                  additional byte of input
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_kyber_shake128_absorb(keccak_state *s, const unsigned char *input, unsigned char x, unsigned char y) {
+    unsigned char extseed[KYBER_SYMBYTES + 2];
+    int i;
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        extseed[i] = input[i];
+    }
+    extseed[i++] = x;
+    extseed[i]   = y;
+    shake128_absorb(s->s, extseed, KYBER_SYMBYTES + 2);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER768_CLEAN_kyber_shake128_squeezeblocks
+*
+* Description: Squeeze step of SHAKE128 XOF. Squeezes full blocks of SHAKE128_RATE bytes each.
+*              Modifies the state. Can be called multiple times to keep squeezing,
+*              i.e., is incremental.
+*
+* Arguments:   - unsigned char *output:      pointer to output blocks
+*              - size_t nblocks:             number of blocks to be squeezed (written to output)
+*              - keccak_state *s:            pointer to in/output Keccak state
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_kyber_shake128_squeezeblocks(unsigned char *output, size_t nblocks, keccak_state *s) {
+    shake128_squeezeblocks(output, nblocks, s->s);
+}
+
+/*************************************************
+* Name:        PQCLEAN_KYBER768_CLEAN_shake256_prf
+*
+* Description: Usage of SHAKE256 as a PRF, concatenates secret and public input
+*              and then generates outlen bytes of SHAKE256 output
+*
+* Arguments:   - unsigned char *output:      pointer to output
+*              - size_t outlen:              number of requested output bytes
+*              - const unsigned char * key:  pointer to the key (of length KYBER_SYMBYTES)
+*              - const unsigned char nonce:  single-byte nonce (public PRF input)
+**************************************************/
+void PQCLEAN_KYBER768_CLEAN_shake256_prf(unsigned char *output, size_t outlen, const unsigned char *key, unsigned char nonce) {
+    unsigned char extkey[KYBER_SYMBYTES + 1];
+    size_t i;
+
+    for (i = 0; i < KYBER_SYMBYTES; i++) {
+        extkey[i] = key[i];
+    }
+    extkey[i] = nonce;
+
+    shake256(output, outlen, extkey, KYBER_SYMBYTES + 1);
+}
diff --git a/crypto_kem/kyber768/clean/symmetric.h b/crypto_kem/kyber768/clean/symmetric.h
new file mode 100644
index 00000000..7cdff295
--- /dev/null
+++ b/crypto_kem/kyber768/clean/symmetric.h
@@ -0,0 +1,26 @@
+#ifndef SYMMETRIC_H
+#define SYMMETRIC_H
+
+#include "fips202.h"
+#include "params.h"
+
+typedef struct {
+    uint64_t s[25];
+} keccak_state;
+
+void PQCLEAN_KYBER768_CLEAN_kyber_shake128_absorb(keccak_state *s, const unsigned char *input, unsigned char x, unsigned char y);
+void PQCLEAN_KYBER768_CLEAN_kyber_shake128_squeezeblocks(unsigned char *output, size_t nblocks, keccak_state *s);
+void PQCLEAN_KYBER768_CLEAN_shake256_prf(unsigned char *output, size_t outlen, const unsigned char *key, unsigned char nonce);
+
+#define hash_h(OUT, IN, INBYTES) sha3_256(OUT, IN, INBYTES)
+#define hash_g(OUT, IN, INBYTES) sha3_512(OUT, IN, INBYTES)
+#define xof_absorb(STATE, IN, X, Y) PQCLEAN_KYBER768_CLEAN_kyber_shake128_absorb(STATE, IN, X, Y)
+#define xof_squeezeblocks(OUT, OUTBLOCKS, STATE) PQCLEAN_KYBER768_CLEAN_kyber_shake128_squeezeblocks(OUT, OUTBLOCKS, STATE)
+#define prf(OUT, OUTBYTES, KEY, NONCE) PQCLEAN_KYBER768_CLEAN_shake256_prf(OUT, OUTBYTES, KEY, NONCE)
+#define kdf(OUT, IN, INBYTES) shake256(OUT, KYBER_SSBYTES, IN, INBYTES)
+
+#define XOF_BLOCKBYTES 168
+
+typedef keccak_state xof_state;
+
+#endif /* SYMMETRIC_H */
diff --git a/crypto_kem/kyber768/clean/verify.c b/crypto_kem/kyber768/clean/verify.c
index bacaa0b7..2fa67412 100644
--- a/crypto_kem/kyber768/clean/verify.c
+++ b/crypto_kem/kyber768/clean/verify.c
@@ -1,45 +1,46 @@
-#include <stdint.h>
-#include <string.h>
-
 #include "verify.h"
 
+#include <stdint.h>
+#include <stdlib.h>
+
 /*************************************************
- * Name:        verify
- *
- * Description: Compare two arrays for equality in constant time.
- *
- * Arguments:   const unsigned char *a: pointer to first byte array
- *              const unsigned char *b: pointer to second byte array
- *              size_t len:             length of the byte arrays
- *
- * Returns 0 if the byte arrays are equal, 1 otherwise
- **************************************************/
-int PQCLEAN_KYBER768_CLEAN_verify(const unsigned char *a, const unsigned char *b, size_t len) {
+* Name:        verify
+*
+* Description: Compare two arrays for equality in constant time.
+*
+* Arguments:   const unsigned char *a: pointer to first byte array
+*              const unsigned char *b: pointer to second byte array
+*              size_t len:             length of the byte arrays
+*
+* Returns 0 if the byte arrays are equal, 1 otherwise
+**************************************************/
+unsigned char PQCLEAN_KYBER768_CLEAN_verify(const unsigned char *a, const unsigned char *b, size_t len) {
     uint64_t r;
     size_t i;
-    r = 0;
 
+    r = 0;
     for (i = 0; i < len; i++) {
         r |= a[i] ^ b[i];
     }
 
-    r = (-(int64_t)r) >> 63;
-    return (int)r;
+    r = (~r + 1); // Two's complement
+    r >>= 63;
+    return (unsigned char)r;
 }
 
 /*************************************************
- * Name:        cmov
- *
- * Description: Copy len bytes from x to r if b is 1;
- *              don't modify x if b is 0. Requires b to be in {0,1};
- *              assumes two's complement representation of negative integers.
- *              Runs in constant time.
- *
- * Arguments:   unsigned char *r:       pointer to output byte array
- *              const unsigned char *x: pointer to input byte array
- *              size_t len:             Amount of bytes to be copied
- *              unsigned char b:        Condition bit; has to be in {0,1}
- **************************************************/
+* Name:        cmov
+*
+* Description: Copy len bytes from x to r if b is 1;
+*              don't modify x if b is 0. Requires b to be in {0,1};
+*              assumes two's complement representation of negative integers.
+*              Runs in constant time.
+*
+* Arguments:   unsigned char *r:       pointer to output byte array
+*              const unsigned char *x: pointer to input byte array
+*              size_t len:             Amount of bytes to be copied
+*              unsigned char b:        Condition bit; has to be in {0,1}
+**************************************************/
 void PQCLEAN_KYBER768_CLEAN_cmov(unsigned char *r, const unsigned char *x, size_t len, unsigned char b) {
     size_t i;
 
diff --git a/crypto_kem/kyber768/clean/verify.h b/crypto_kem/kyber768/clean/verify.h
index 8a1b81f4..ecfe79a9 100644
--- a/crypto_kem/kyber768/clean/verify.h
+++ b/crypto_kem/kyber768/clean/verify.h
@@ -3,7 +3,7 @@
 
 #include <stdio.h>
 
-int PQCLEAN_KYBER768_CLEAN_verify(const unsigned char *a, const unsigned char *b, size_t len);
+unsigned char PQCLEAN_KYBER768_CLEAN_verify(const unsigned char *a, const unsigned char *b, size_t len);
 
 void PQCLEAN_KYBER768_CLEAN_cmov(unsigned char *r, const unsigned char *x, size_t len, unsigned char b);
 
diff --git a/test/test_nistkat.py b/test/test_nistkat.py
index 97c8293f..daebca49 100644
--- a/test/test_nistkat.py
+++ b/test/test_nistkat.py
@@ -25,10 +25,6 @@ def test_nistkat():
 
 @helpers.filtered_test
 def check_nistkat(implementation):
-    if implementation.scheme.name == "kyber768":
-        raise unittest.SkipTest(
-            "Temporarily skip NIST KAT check for kyber768 since it's "
-            "an outdated implementation")
     helpers.make('nistkat',
                  TYPE=implementation.scheme.type,
                  SCHEME=implementation.scheme.name,