Mistral Small 4

Mistral Small 4 represents a consolidation of three distinct Mistral model lineages—Instruct, Reasoning (previously Magistral), and Devstral—into a single hybrid system. Rather than being a simple upgrade, it functions as both a general instruction model and a reasoning model, with mode switching configurable per request via a reasoning effort parameter that toggles between fast instant replies and deeper test-time compute. This unification aims to give developers a single versatile endpoint for diverse tasks rather than routing between specialized models.

The architecture is a mixture-of-experts design with 119 billion total parameters but only 6.5B activated per token across 128 experts (4 active at a time), supporting a long context window for substantial inputs. Multimodal capability allows the model to analyze images alongside text, and it supports dozens of languages. Compared to its predecessor, Mistral Small 4 claims roughly a 40% reduction in end-to-end completion latency in latency-optimized setups and approximately 3x higher request throughput, with optional speculative decoding and NVFP4 quantization checkpoints available for further efficiency gains. Its combination of reasoning, vision, and agentic coding strengths makes it well suited for general-purpose assistants, document and image analysis workflows, and developer-oriented tooling where flexible depth of inference is valuable.