Mistral Small 3.1 24B

Mistral Small 3.1 24B is an instruction-finetuned model built from Mistral-Small-3.1-24B-Base-2503, which itself extends the Mistral Small 3 2501 lineage. The 24-billion-parameter release is positioned as a multimodal step forward, adding state-of-the-art vision understanding alongside its existing text and code strengths without regressing on standard language benchmarks. Long-context handling is extended up to 128k tokens, making it well suited to workloads that combine document-length reasoning with visual inputs such as diagrams, screenshots, or annotated figures.

For practical fit, the model is aimed at teams that need a compact open-weight instruct model with both language and vision support, including conversational assistants, document analysis, and structured output generation. Mistral AI documents deployment guidance recommending vLLM 0.9.1 or newer (with the bundled mistral_common 1.6.2), low sampling temperatures around 0.15, and use of the shipped SYSTEM_PROMPT.txt to shape behavior. A later successor, Mistral-Small-3.2-24B-Instruct-2506, is described by Mistral as a minor refinement of this 2503 Instruct release, indicating an active iteration path while the underlying strengths of the 3.1 line remain the foundation for current usage.