Mistral Small 4: 119B MoE Apache 2.0 Model Unifies Reasoning, Vision, and Code with Per-Request reasoning_effort Control

simonwillison.net·high signal

Mistral released Mistral Small 4, a 119B parameter Mixture-of-Experts model with only 6B active parameters per token, licensed Apache 2.0. The model introduces a reasoning_effort parameter (none/high) giving developers direct cost-vs-depth control per request, and delivers 40% faster completions and 3x more requests/second than its predecessor. Available day-0 on NVIDIA NIM, vLLM, llama.cpp, and HuggingFace — the first single open checkpoint unifying vision, code agents, and configurable chain-of-thought reasoning under a permissive license.