Mistral Small 4 is designed as a unified hybrid model that consolidates what were once separate lineages into a single system, combining general instruction capabilities, reasoning features previously branded as Magistral, and the agentic coding strengths of Devstral. Built on a mixture-of-experts architecture with 128 experts and roughly 6.5 billion parameters activated per token out of a total 119 billion, the model is engineered to switch between fast instruction-following and deeper reasoning on a per-request basis, giving developers one checkpoint that can serve many task profiles. Its multimodal front end accepts both text and image inputs while returning text, making it suitable for workflows that mix documents, screenshots, or diagrams with conversational or analytical prompts.

In practice, Mistral Small 4 targets a broad sweet spot of everyday enterprise and developer use, from coding assistance and tool-driven agents to multilingual chat and visual question answering. The model is delivered as an open-weight release with a long context window, and it exposes a configurable reasoning effort knob for tuning latency versus answer quality. Compared with the prior Mistral Small generation, reported gains include a 40 percent reduction in end-to-end completion time in latency-optimized serving and roughly three times the request throughput in throughput-optimized deployments, alongside optional efficiency paths such as speculative decoding with a trained eagle head and NVFP4 quantization. These properties make it a flexible choice for teams that want one generalist system to cover reasoning, vision, and agentic work without juggling multiple specialized models.