Mistral Small 4 is the new __hybrid model__ from Mistral AI, merging Magistral capabilities for __reasoning__, Pixtral for __multimodal__, and Devstral for __code agentic__. __Mixture of Experts__ architecture with 128 experts of which 4 are active per token, 119 billion total parameters for 6 billion active. Distributed under __Apache 2.0 license__, it integrates via Mistral API, La Plateforme, and Le Chat, and powers both conversational assistants and autonomous __developer agents__.
What is Mistral Small 4?
Mistral Small 4 is an open source hybrid language model designed by Mistral AI. Its architecture relies on a Mixture of Experts with 128 experts of which 4 are activated per token, for a total of 119 billion parameters with only 6 billion active at inference. This approach brings significant energy and economic efficiency without sacrificing model depth. The model natively accepts text and image inputs, supports configurable reasoning effort, and excels particularly at agentic and code tasks. The Apache 2.0 license allows downloading, modifying, and deploying it, including for commercial uses.
Key Features
Mistral Small 4 combines several technical advances. Its Mixture of Experts architecture allows scaling total parameters without penalizing inference cost, making the model more accessible to deploy on reasonably-sized infrastructure. Native multimodal capability removes the need for an additional vision model to process text and images together. Configurable reasoning effort allows trading off between speed and analysis depth depending on use case. Specialization in code agentic tasks, inherited from Devstral, makes it an excellent engine for developer copilots and autonomous agents. The model is available via Mistral Chat, La Plateforme, and the official API, but is also downloadable for on-premise deployments. Mistral AI joined the NVIDIA Nemotron Coalition as founding member, securing the GPU optimization ecosystem around the model.
Use Cases
Mistral Small 4 use cases span a wide spectrum. Developers integrate it into code copilots and autonomous agents capable of executing complex technical tasks. Data science teams use it for structured reasoning, document analysis, and information extraction on large corpora. Customer service deploys conversational assistants capable of understanding screenshots and scanned documents thanks to multimodality. Sovereign organizations, particularly in public, financial, or healthcare sectors, adopt it to benefit from AI deployable on-premise under open license. Startups build their products on the Mistral API layer to combine controlled cost and model quality. Researchers appreciate the model's accessibility for evaluating new alignment methods, fine-tuning, and quantization approaches.
Benefits
Mistral Small 4's main benefit lies in converging three model families into one. This unification simplifies AI application architecture by eliminating the need for routing between multiple specialized models. The Apache 2.0 license brings major strategic independence for organizations unwilling to depend on proprietary APIs. MoE architecture efficiency enables considering cost-controlled deployments while benefiting from high quality. Maturity of the Mistral ecosystem, with its Plateforme, API, and Chat, offers multiple entry points depending on team technical level. Finally, the European origin of the model is a decisive argument for actors sensitive to GDPR compliance and digital sovereignty.
Pricing
Mistral Small 4 is made available free for download under Apache 2.0 license, allowing on-premise deployment without license fees. Access via Mistral API is billed on usage based on tokens consumed, at competitive rates aligned with the model's efficiency strategy. Mistral Chat offers free access with quotas to test the model's capabilities, plus paid plans for intensive use. La Plateforme includes monitoring and fine-tuning tools for professional teams. Overall, Mistral Small 4 offers excellent pricing flexibility, from free on-premise deployment to serverless API usage, through enterprise plans for high-volume organizations.
Conclusion
Mistral Small 4 confirms Mistral AI's strategy: deliver open, performant, and technically differentiated models. Its multimodal versatility and configurable reasoning effort make it suitable for many use cases, from developer copilot to multimodal assistant. The Apache 2.0 license and on-premise deployment possibility will particularly appeal to European players and sovereign organizations. For technical teams seeking a flexible, high-performing, and independent model from American giants, Mistral Small 4 is today a major reference in the open source AI landscape.