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 */
/*
* sock.h - the few things a TCP socket does differently on Windows: it is
* not a file descriptor there (recv and send, not read and write; closed
* with closesocket), its errors come from WSAGetLastError, and Winsock
* must be started before any of it, name lookups included. These calls
* leave the error in errno, with POSIX's names, on every system.
*/
#ifndef JANAS_COMMON_SOCK_H
#define JANAS_COMMON_SOCK_H
#include <stddef.h>
/* Winsock started, once (nothing elsewhere). Before getaddrinfo too. */
void janas_sock_init(void);
/* A socket that does not block and is not passed on to programs started
later, or -1. */
int janas_sock_open(int family, int type, int protocol);
/* Whether the connect() that just failed is only under way (EINPROGRESS;
Windows: WSAEWOULDBLOCK). */
int janas_sock_connecting(void);
long janas_sock_recv(int fd, void *p, size_t n);
long janas_sock_send(int fd, const void *p, size_t n);
int janas_sock_close(int fd);
#endif