v0.1.2

CPU inference for Qwen3.5-0.8B on Windows and Linux, built from ef49a53da08949f536408177b1ee78a034baf601.

• Faster H128 prefill on supported AVX-512/VNNI CPUs, with runtime ISA dispatch. The measured Ryzen workload improves prefill by approximately 26–29%; results depend on CPU and workload.
• Download scripts pin the current H128 B+C 256 calibrated checkpoint (f6cb5cf04a9094670f9041578a3f4e44b94a1395). Model weights are downloaded separately.
• Native Q4_0 GGUF execution and updated quantization/evaluation tools.
• Current speed and quality comparisons against llama.cpp and ik_llama.cpp, including batch-16 decode.

Download the Windows x64 ZIP or Linux x64 tar.gz below and extract it. Each archive includes the CLI, HTTP server, model download scripts and usage instructions. Windows uses a static C++ runtime; Linux targets Ubuntu 22.04 or newer (glibc 2.35+). No compiler or Python is required to run the packaged engine.

Both platform packages pass kernel tests, real-model scheduler tests and extracted-package HTTP verification before publication. SHA-256 checksum files accompany both archives.

Usage and current benchmarks · Full changelog

v0.1.1

Full Changelog: v0.1.0...v0.1.1

v0.1.0

Full Changelog: https://github.com/Danmoreng/qwen35-cpu/commits/v0.1.0