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 2026-10-12 (3): the cores' efficiency class read with older MinGW-w64 headers too
camauri06ec016d620d
diff --git a/src/common/sysinfo_win.c b/src/common/sysinfo_win.cindex a182082..65f3e8a 100644--- a/src/common/sysinfo_win.c+++ b/src/common/sysinfo_win.c@@ -59,7 +59,9 @@ static int cores_of(struct core *out, int max)struct core *c = &out[n++];c->n_cpus = 0;c->first = -1;- c->cls = e->Processor.EfficiencyClass;+ /* the byte after Flags: EfficiencyClass in the headers of+ MinGW-w64 12 and later, Reserved[0] in the older ones */+ c->cls = ((const BYTE *)&e->Processor)[1];for (WORD g = 0; g < e->Processor.GroupCount; g++) {GROUP_AFFINITY *ga = &e->Processor.GroupMask[g];for (int b = 0; b < 64; b++)