Mistral Small 4

Mistral Small 4 is designed as a single hybrid model that brings together the strengths of three previously separate Mistral lines: a general Instruct model, the Reasoning family (formerly known as Magistral), and Devstral for coding. Instead of choosing between a chat model and a reasoning model, users can stay on one checkpoint and toggle between fast instant replies and a deeper reasoning mode whose effort is configurable per request. The result is a general-purpose model aimed at assistants, agents, and developer workflows that need both conversational fluency and structured problem solving without switching backends.

Under the hood, Mistral Small 4 uses a sparse Mixture-of-Experts design with 119 billion total parameters but only 6.5 billion active per token, spread across 128 experts with 4 active at a time. That sparsity is paired with a long context window and multimodal inputs that accept both text and images while producing text outputs, making the model suitable for document, screenshot, and code-scene understanding. Mistral also reports roughly a 40 percent reduction in end-to-end completion latency and about three times the throughput of Mistral Small 3 in optimized serving setups, and ships companion artifacts for further speed-ups, including a trained Eagle speculative-decoding head and an NVFP4 4-bit quantization checkpoint, giving operators flexible paths to fit different latency and cost budgets.