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 */
/*
* layout.h - the layouts of the services' answers in janas-chat
* (services/common/template.h): an answer that brings its data and an English
* layout is shown to the user filled in the user's language, and the model
* reads only its brief. The layout in another language is asked of the
* model once, checked against the English one and kept in
* ~/.config/janas/layouts; the scales' terms come from the glossary
* (glossary.h), not from the model.
*/
#ifndef JANAS_CHAT_LAYOUT_H
#define JANAS_CHAT_LAYOUT_H
#include <stddef.h>
/* The user's language ("it"; "en" or NULL: English, no translation). */
void layout_set_lang(const char *lang);
/* The model the layouts are translated by: each model keeps its own, and
its failures, so that a small model's poor translation is never shown
with another model's answers and a failed one is not asked for again. */
void layout_set_model(const char *key);
const char *layout_lang(void);
/* How a layout is translated: system and user messages to the model, its
answer into *out (malloc'd). 0, or -1. Called once a layout and
language; what it says on the terminal is its own. */
void layout_set_translator(int (*fn)(const char *name, const char *system,
const char *text, char **out));
/* A tools/call result (JSON) with a layout: 1 with the text for the user
in *user and the brief for the model in *model (both malloc'd); 0 when
it has none, or it could not be filled (then the answer is used as it
came). */
int layout_take(const char *result, size_t n, char **user, char **model);
#endif