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 */
/*
* steps.h - what janas-get does to a file: download it and check it, run a
* converter beside janas-get, check what it wrote.
*/
#ifndef JANAS_GET_STEPS_H
#define JANAS_GET_STEPS_H
#include <stddef.h>
/* dir/the last part of path */
void get_local(char *out, size_t len, const char *dir, const char *path);
/* File path of repo downloaded into dir (resumed if part of it is there),
its SHA-256 checked against want (NULL: the one Hugging Face lists). 0,
-1 having said why, or -2 when Hugging Face does not list the file
(nothing written: the caller may take another way, quiet says nothing
then). */
int get_fetch(const char *repo, const char *path, const char *dir,
const char *want, int quiet);
/* 1 when file's SHA-256 is want; 0 when it is another, -1 unreadable. */
int get_same(const char *file, const char *want);
/* A converter of the same build (gguf2jns, jns_planes, hf2jns_mtp) run on
argv (argv[0] its name, NULL-terminated). Its exit status, or -1. */
int get_run(const char *tool, char *const argv[]);
#endif