Mistral Releases Small 4: 119B MoE That Unifies Instruct, Reasoning, Vision, and Coding in One Model

Mistral AI·high signal

Mistral AI released Mistral Small 4 on March 16 — a 119B-parameter Mixture-of-Experts model with only 6B active parameters per token and a 256k context window that consolidates Mistral Small, Magistral (reasoning), Pixtral (vision), and Devstral (agentic coding) into a single deployment target. Benchmarks show 40% faster end-to-end completion and 3x higher throughput than Small 3. Released under Apache 2.0, it runs on vLLM and is available on Hugging Face, making it the most capable open-weight model that requires no model switching across multimodal and agentic workflows.