Mistral Small 4 Released

Mistral Small 4 is Mistral's new flagship model that integrates the reasoning of Magistral, the multimodality of Pixtral, and the agentic coding ability of Devstral into a single model. It is released under the Apache 2.0 license, allowing a single model to cover what previously required switching between a fast instruct model, a strong reasoning engine, and a multimodal assistant.

The model is designed with a hybrid architecture that supports both text and image inputs, and uses a Mixture of Experts(MoE) architecture. Out of 128 experts in total, only 4 experts are activated per token, with 119B total parameters and 6B active parameters per token (8B when including embedding and output layers), and a 256k context window that handles long documents and lengthy conversations.

The core feature is controlled via the reasoning_effort parameter. reasoning_effort="none" provides fast, lightweight responses at the level of Mistral Small 3.2, while reasoning_effort="high" provides step-by-step reasoning at a depth similar to the previous Magistral model.

In terms of performance, internal comparisons show a 40% reduction in end-to-end completion time and a 3x improvement in requests per second. Even with reasoning turned on, the model produced competitive scores across three benchmarks compared to GPT-OSS 120B while generating shorter outputs, and on AA LCR it scored 0.72 with 1.6K characters, achieving similar performance with far fewer output characters than the Qwen series. On LiveCodeBench, it outperformed GPT-OSS 120B while using 20% less output.

From a deployment perspective, the minimum configuration presented is 4x NVIDIA HGX H100, 2x NVIDIA HGX H200, or 1x NVIDIA DGX B200, while the recommended configuration is 4x HGX H100, 4x HGX H200, or 2x DGX B200. The model can be used with vLLM, llama.cpp, SGLang, Transformers, and more, and day-0 deployment is also available via NVIDIA NIM.

Key use cases are as follows.

• Developers: code automation, codebase exploration, agentic coding workflows
• Enterprises: general chat assistants, document understanding, multimodal analysis
• Researchers: math, research, complex reasoning

The core message is clear. By combining instruct, reasoning, and multimodal capabilities into one, Mistral Small 4 is a model that aims for both the flexibility of an open-source foundation and enterprise-grade efficiency at the same time.