37 tok/s from Qwen3-Next-80B (48 GB) on a 32 GB laptop, 24 without drafts vs 2.8 for llama.cpp; reading 62 vs 16 tok/s. Faster than llama.cpp on all 13 models compared in Q4_K_M. LLMs bigger than your RAM, pure C, CPU or GPU.
C avx2ccpu-inferenceggufinference-enginelinuxllmlocal-llmmixture-of-expertsmoenvmequantizationqwen3simdspeculative-decodingvulkan
/* SPDX-License-Identifier: GPL-3.0-or-later
Copyright (C) 2026 Maurizio Cammalleri */
/*
* md5.c - MD5 as RFC 1321 gives it (see md5.h).
*/
#include "common/md5.h"
#include <string.h>
static const uint32_t K[64] = {
0xd76aa478, 0xe8c7b756, 0x242070db, 0xc1bdceee, 0xf57c0faf, 0x4787c62a,
0xa8304613, 0xfd469501, 0x698098d8, 0x8b44f7af, 0xffff5bb1, 0x895cd7be,
0x6b901122, 0xfd987193, 0xa679438e, 0x49b40821, 0xf61e2562, 0xc040b340,
0x265e5a51, 0xe9b6c7aa, 0xd62f105d, 0x02441453, 0xd8a1e681, 0xe7d3fbc8,
0x21e1cde6, 0xc33707d6, 0xf4d50d87, 0x455a14ed, 0xa9e3e905, 0xfcefa3f8,
0x676f02d9, 0x8d2a4c8a, 0xfffa3942, 0x8771f681, 0x6d9d6122, 0xfde5380c,
0xa4beea44, 0x4bdecfa9, 0xf6bb4b60, 0xbebfbc70, 0x289b7ec6, 0xeaa127fa,
0xd4ef3085, 0x04881d05, 0xd9d4d039, 0xe6db99e5, 0x1fa27cf8, 0xc4ac5665,
0xf4292244, 0x432aff97, 0xab9423a7, 0xfc93a039, 0x655b59c3, 0x8f0ccc92,
0xffeff47d, 0x85845dd1, 0x6fa87e4f, 0xfe2ce6e0, 0xa3014314, 0x4e0811a1,
0xf7537e82, 0xbd3af235, 0x2ad7d2bb, 0xeb86d391};
static const uint8_t R[64] = {
7, 12, 17, 22, 7, 12, 17, 22, 7, 12, 17, 22, 7, 12, 17, 22,
5, 9, 14, 20, 5, 9, 14, 20, 5, 9, 14, 20, 5, 9, 14, 20,
4, 11, 16, 23, 4, 11, 16, 23, 4, 11, 16, 23, 4, 11, 16, 23,
6, 10, 15, 21, 6, 10, 15, 21, 6, 10, 15, 21, 6, 10, 15, 21};
static void block(uint32_t h[4], const uint8_t *p)
{
uint32_t m[16];
for (int i = 0; i < 16; i++)
m[i] = (uint32_t)p[4 * i] | (uint32_t)p[4 * i + 1] << 8 |
(uint32_t)p[4 * i + 2] << 16 | (uint32_t)p[4 * i + 3] << 24;
uint32_t a = h[0], b = h[1], c = h[2], d = h[3];
for (int i = 0; i < 64; i++) {
uint32_t f;
int g;
if (i < 16) {
f = (b & c) | (~b & d);
g = i;
} else if (i < 32) {
f = (d & b) | (~d & c);
g = (5 * i + 1) % 16;
} else if (i < 48) {
f = b ^ c ^ d;
g = (3 * i + 5) % 16;
} else {
f = c ^ (b | ~d);
g = (7 * i) % 16;
}
uint32_t t = d;
d = c;
c = b;
uint32_t x = a + f + K[i] + m[g];
b = b + (x << R[i] | x >> (32 - R[i]));
a = t;
}
h[0] += a;
h[1] += b;
h[2] += c;
h[3] += d;
}
void janas_md5_init(struct janas_md5 *s)
{
s->h[0] = 0x67452301;
s->h[1] = 0xefcdab89;
s->h[2] = 0x98badcfe;
s->h[3] = 0x10325476;
s->len = 0;
s->n = 0;
}
void janas_md5_update(struct janas_md5 *s, const void *data, size_t n)
{
const uint8_t *p = data;
s->len += n;
if (s->n) {
size_t k = 64 - s->n < n ? 64 - s->n : n;
memcpy(s->buf + s->n, p, k);
s->n += k;
p += k;
n -= k;
if (s->n < 64)
return;
block(s->h, s->buf);
s->n = 0;
}
for (; n >= 64; p += 64, n -= 64)
block(s->h, p);
memcpy(s->buf, p, n);
s->n = n;
}
void janas_md5_hex(struct janas_md5 *s, char out[33])
{
uint64_t bits = s->len * 8;
uint8_t pad[72] = {0x80};
size_t k = s->n < 56 ? 56 - s->n : 120 - s->n;
for (int i = 0; i < 8; i++)
pad[k + (size_t)i] = (uint8_t)(bits >> (8 * i));
janas_md5_update(s, pad, k + 8);
static const char hex[] = "0123456789abcdef";
for (int i = 0; i < 16; i++) {
uint8_t b = (uint8_t)(s->h[i / 4] >> (8 * (i % 4)));
out[2 * i] = hex[b >> 4];
out[2 * i + 1] = hex[b & 15];
}
out[32] = 0;
}