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 */
/*
* win_start.c - what every program of Janas does first on Windows, before
* main: the console made to show UTF-8, and to take the escape sequences of
* colours and of the cursor as a Linux terminal does. Linked into the
* programs only (build.sh), never into the libraries: a program of
* someone else's that loads them keeps its console as it set it.
*/
#ifdef _WIN32
#include <windows.h>
__attribute__((constructor)) static void janas_win_start(void)
{
SetConsoleOutputCP(CP_UTF8);
DWORD std[2] = {STD_OUTPUT_HANDLE, STD_ERROR_HANDLE};
for (int i = 0; i < 2; i++) {
HANDLE h = GetStdHandle(std[i]);
DWORD m;
if (h != INVALID_HANDLE_VALUE && GetConsoleMode(h, &m))
SetConsoleMode(h, m | ENABLE_VIRTUAL_TERMINAL_PROCESSING);
}
}
#endif