Mistral releases Small 4, a 119-billion-parameter MoE model for reasoning, code and multimodal work

Mistral AI has introduced Mistral Small 4, a new open model meant to replace several of the company's separate product lines with a single, universal endpoint. Instead of a separate instruct model, reasoning model, vision model and coding agent, developers get one MoE checkpoint with adjustable reasoning depth.

One instead of four

The point of the release is not that Mistral simply added more parameters. Small 4 merges into one product the roles previously split among Mistral Small for ordinary instructions, Magistral for complex reasoning, Pixtral for multimodal understanding and Devstral for agentic programming. For teams building products on top of LLMs, that matters more than another record in a benchmark table: less routing between models, simpler infrastructure, and a lower chance of getting different answer styles on neighboring steps of the same workflow.

"Users no longer need to choose between a fast instruct mode, reasoning and a multimodal assistant,"

Mistral says in its announcement.

By its positioning, Small 4 targets several kinds of work at once: regular chat, coding, agentic workflows, and analysis of complex documents or images. Mistral pitches the model outright as a universal layer for enterprise tasks that need text and visual queries under a single API surface. That stands out in a market where many teams still run separate models for chat, for reasoning and for vision.

How the model is built

Architecturally, it is a Mixture-of-Experts model with 119 billion parameters. It has 128 experts, of which only four activate for each token, so Mistral is betting on efficiency at inference rather than maximum density. The company also states a 256k context window and native handling of text and images. The release is open under the Apache 2.0 license, which means the model can be used through the API as well as deployed and fine-tuned for your own scenarios.

• 119 billion parameters in the overall architecture
• 128 experts and 4 active experts per token
• 256k context window
• Inputs: text and images
• Apache 2.0 license and availability for self-hosting

Mistral puts particular emphasis on the reasoning_effort parameter. In essence, it switches between a fast answer and a heavier step-by-step reasoning mode. With none, the model is supposed to behave closer to Mistral Small 3.2 and return lighter, low-latency answers. With high, it should work closer to the Magistral line, when reasoning quality on hard tasks matters more than speed. The practical upshot is simple: instead of a stack of two or three models, you can run one deployment and change behavior per request.

Speed and deployment

In its official announcement, Mistral bets not only on versatility but also on inference economics. The company claims a 40% reduction in end-to-end completion time in a latency-optimized configuration, and a threefold increase in requests per second in a throughput-optimized setup, compared with Mistral Small 3. Mistral also stresses that Small 4 with reasoning enabled delivers comparable or better results than GPT-OSS 120B on AA LCR, LiveCodeBench and AIME 2025, while generating shorter outputs. These comparisons were published by the company itself, but the focus on "quality per token" for production really does matter.

Mistral also lists practical deployment options up front. The model is available through the Mistral API and AI Studio, is posted on Hugging Face, and is declared compatible with vLLM, llama.cpp, SGLang and Transformers. Self-hosting requirements are no longer "desktop": the minimum configuration is 4x NVIDIA HGX H100, 2x HGX H200 or 1x DGX B200, and more powerful setups are recommended for best performance. In other words, Small 4 looks like an open model not for a laptop but for serious server infrastructure and product teams that value control, customization and predictable cost of ownership.

What it means

Mistral is pushing the open-source segment toward more universal models, where the main advantage is not only quality but also a simpler system around the LLM. If Small 4 confirms the claimed efficiency in real production workloads, the company will have a strong argument against a zoo of separate reasoning, vision and coding models. For businesses, it is a chance to cut the complexity of the orchestration layer; for developers, a chance to get one configurable base layer for a wide range of tasks.