Mistral Small 4 is presented by third-party cataloging as the next major release in the Mistral Small family, unifying capabilities from several flagship Mistral models into a single system with an emphasis on strong reasoning. Community deployment write-ups characterize it as a 119B-parameter mixture-of-experts architecture that can be served locally using SGLang on DGX Spark (GB10) hardware, suggesting it is designed to scale across both hosted and self-hosted inference environments. This combination of open weights, a substantial expert capacity, and community-validated serving paths positions it as a flexible foundation for teams that want to move between API use and local deployment.

In practical terms, Mistral Small 4 is framed for workflows that blend text and image understanding with structured tool use. Catalog listings highlight it as well suited for image understanding, long documents, tool-augmented tasks, structured responses, and deep analysis, with reported catalog index scores of 26.6 for coding, 4.6 for agentic use, and 19.7 for intelligence, alongside a Krater rank of 59. A context length of roughly 262K tokens with a maximum output around 210K tokens supports long-form and document-heavy applications, while native support for tool calling and structured outputs makes it a practical choice for agent pipelines and reproducible integrations where reasoning quality and multimodal input both matter.