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
Release 2026-10-11 (3): the build on older C libraries
camaurib90858a2fc1e
diff --git a/build.sh b/build.shindex c36cf70..091a111 100755--- a/build.sh+++ b/build.sh@@ -99,7 +99,10 @@ BIN="bin/$CPU-$OS"SUFFIX=""[ "$VARIANT" != release ] && SUFFIX="-$VARIANT"-CFLAGS=(-std=gnu11 -Wall -Wextra -Werror -pthread -fPIC -fvisibility=hidden+# _GNU_SOURCE for every file, empty as the files that define it write it:+# the C library of Ubuntu 22.04 (the binary package's) declares memmem,+# strcasestr and the like only with it, where a newer one does by default+CFLAGS=(-std=gnu11 -D_GNU_SOURCE= -Wall -Wextra -Werror -pthread -fPIC -fvisibility=hidden-Isrc -Iinclude)case "$VARIANT" inrelease) CFLAGS+=(-O2 -DNDEBUG) ;;diff --git a/tests/test_github_parse.c b/tests/test_github_parse.cindex ea22a7c..286bc07 100644--- a/tests/test_github_parse.c+++ b/tests/test_github_parse.c@@ -10,7 +10,6 @@* directory of its own under ~/tmp.* The sources belong to the program, so they are compiled in here.*/-#define _GNU_SOURCE /* memmem: Ubuntu 22.04's glibc (2.35) declares it only so */#include <stdio.h>#include <stdlib.h>#include <string.h>