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 */
/*
* env_names.h - the variables of the environment Janas reads (env_names.c),
* so that one set with a name Janas does not know - JANAS_EXPERT_BIT for
* JANAS_EXPERT_BITS - is told when a program starts, instead of doing
* nothing in silence. A warning, not a refusal: a variable may be meant for
* another version of Janas.
*/
#ifndef JANAS_COMMON_ENV_NAMES_H
#define JANAS_COMMON_ENV_NAMES_H
#include <stddef.h>
/* every JANAS_ variable read, by the programs, the library or the
services; tests/test_env_names.c checks it against the sources */
extern const char *const JANAS_ENV_NAMES[];
extern const size_t JANAS_N_ENV_NAMES;
/* On stderr, prefixed by prog, a line for each JANAS_ variable of the
environment not among them, with the name nearest to it. The number of
them. */
int janas_env_check(const char *prog);
#endif