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 */
/*
* image.c - a picture's format told by its first bytes, and the decoder
* of that format (image.h).
*/
#include "image.h"
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int janas_image_decode(const void *data, size_t n, struct janas_image *im,
char *why, size_t wn)
{
const uint8_t *p = data;
memset(im, 0, sizeof(*im));
if (n >= 8 && !memcmp(p, "\x89PNG\r\n\x1a\n", 8))
return janas_png_decode(p, n, im, why, wn);
if (n >= 3 && p[0] == 0xff && p[1] == 0xd8 && p[2] == 0xff)
return janas_jpeg_decode(p, n, im, why, wn);
if (n >= 12 && !memcmp(p, "RIFF", 4) && !memcmp(p + 8, "WEBP", 4))
snprintf(why, wn, "WebP pictures are not read yet: send PNG or JPEG");
else
snprintf(why, wn, "not a picture this reads (PNG or JPEG)");
return -1;
}
void janas_image_free(struct janas_image *im)
{
free(im->rgb);
memset(im, 0, sizeof(*im));
}