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 (2): the Windows build's packages installed with no question asked
camauri69f47d98fbdb
diff --git a/.github/workflows/release.yml b/.github/workflows/release.ymlindex 3b72f69..3f3cc9a 100644--- a/.github/workflows/release.yml+++ b/.github/workflows/release.yml@@ -125,14 +125,18 @@ jobs:# headers do), then tested on Windows, then packed.windows-build:runs-on: ubuntu-24.04+ timeout-minutes: 40steps:- uses: actions/checkout@v7- name: MinGW-w64, the shaders' compiler and the Vulkan headersrun: |set -euo pipefail- sudo apt-get update -q- sudo apt-get install -y -q gcc-mingw-w64-x86-64 binutils-mingw-w64-x86-64+ # no question may wait for an answer (needrestart asks one)+ export DEBIAN_FRONTEND=noninteractive NEEDRESTART_MODE=a+ sudo -E apt-get update -q+ sudo -E apt-get install -y -q --no-install-recommends \+ gcc-mingw-w64-x86-64 binutils-mingw-w64-x86-64mkdir -p "$RUNNER_TEMP/khronos"cd "$RUNNER_TEMP/khronos"curl -fsSL -o glslang.tar.gz \@@ -168,6 +172,7 @@ jobs:windows-test:needs: windows-buildruns-on: windows-latest+ timeout-minutes: 40steps:- uses: actions/checkout@v7@@ -190,6 +195,7 @@ jobs:windows-package:needs: [linux-x86_64, windows-build, windows-test]runs-on: ubuntu-24.04+ timeout-minutes: 20env:WPACKAGE: janas-windows-x86_64steps: