mistral.rs v0.9.0 achieves up to 1.8x faster CPU decode than llama.cpp

The mistral.rs project released version 0.9.0, delivering significant performance improvements for CPU-based inference across x86 and ARM architectures. Benchmarks on Qwen3 4B Q4_K models demonstrate that mistral.rs decodes faster than llama.cpp at every measured context depth.

• The update includes granular optimizations designed to provide general speedups for all supported models.
• Performance gains apply to CPUs of various capabilities, including x86 processors with AVX2 or AVX512 and ARM processors with NEON.
• Testing was conducted by sweeping configurations for both engines to ensure each point represented the best possible performance for that specific engine.

The release allows users to run models like Qwen 3.5/3.6, Gemma 4, and LFM 2.5 directly from Hugging Face using the mistral.rs ISQ system.