Mistral Small 4: open-weights hybrid reasoning

Mistral released a 119B mixture-of-experts model with just 6.5B active parameters per token, supporting both reasoning and non-reasoning modes. Open-weights models continue closing the gap with proprietary ones.