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 */
/*
* request.h - what request.c and request_opts.c share: the reading of a
* request's JSON into a job. Private to them.
*/
#ifndef JANAS_SERVER_REQUEST_H
#define JANAS_SERVER_REQUEST_H
#include <yyjson.h>
#include "server.h"
/* The error of a request that cannot be served, for the 400 answer: always
returns 0, so that BAD() can sit inside an expression. */
int srv_perr(struct srv_perr *e, const char *code, const char *fmt, ...)
__attribute__((format(gnu_printf, 3, 4)));
#define BAD(...) srv_perr(e, __VA_ARGS__)
/* Marks the request failed; returns -1. */
int srv_pfail(struct srv_parsed *pr, int ret);
char *srv_dup_n(const char *s, size_t n);
/* A number between lo and hi: 1 with *out set, 0 absent (or null), -1 not a
number or out of range. */
int srv_num(yyjson_val *v, double lo, double hi, double *out);
/* A text part's text ("refusal" for a refusal, else "text"), NULL if absent;
it may be a value of any kind, the caller checks it is a string. */
yyjson_val *srv_part_text(yyjson_val *part, const char *type);
/* A ceiling on the reasoning by llama.cpp's names, reasoning_budget_tokens
or thinking_budget_tokens: -1 none, 0 no reasoning, N at most N tokens of
it, then the answer. 0, or -1 when the value is not one of those. */
int srv_reasoning_budget(yyjson_val *root, struct janas_llm_chat_params *p);
/* A value as JSON text (NULL on memory). */
char *srv_json_text(yyjson_val *v, size_t *n);
/* request_opts.c: the parameters of both endpoints, and those only chat
has (tools, response_format, logprobs...). */
int srv_read_params(struct srv_perr *e, struct srv_parsed *pr, yyjson_val *root,
struct srv_job *j, int chat);
int srv_read_chat_opts(struct srv_perr *e, struct srv_parsed *pr,
yyjson_val *root, struct srv_job *j);
#endif