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 */
/*
* serve.h - janas-mcp, Janas as an MCP server: what main.c (the protocol)
* and tools.c (the tools) share.
*/
#ifndef JANAS_MCPD_SERVE_H
#define JANAS_MCPD_SERVE_H
#include <stdio.h>
#include "common/mcp_server.h"
#include "janas/llm.h"
#include "llm/json.h"
struct mcpd {
janas_llm *llm; /* the model that writes (NULL: none) */
janas_llm_chat *chat;
char name[128];
struct janas_llm_chat_params cp; /* as started; a call may change some */
janas_llm *emb; /* the model that gives embeddings (NULL: none) */
char emb_name[128];
int call_seconds; /* longest a generation may run (0: no limit) */
struct janas_mcps srv; /* the protocol (common/mcp_server.c) */
};
/* While a call runs: what the client sent meanwhile, taken in. 1 when the
call running was cancelled. */
static inline int mcpd_poll(struct mcpd *d)
{
return janas_mcps_poll(&d->srv);
}
/* tools.c: the tools/list members ("tools": [...]); a tools/call into the
members of its result (0), or -1 for a tool there is not. */
void mcpd_tools_list(const struct mcpd *d, struct janas_buf *b);
int mcpd_tools_call(struct mcpd *d, const struct janas_json *params,
struct janas_buf *b);
#endif