Mistral Small 3.1

Mistral Small 3.1 is a compact, multimodal large language model positioned as a versatile generalist that pairs text understanding with vision capabilities in a single 24-billion-parameter package. Building directly on the earlier Mistral Small 3, it introduces state-of-the-art image understanding while extending the context window far beyond typical small-model limits, aiming to handle long documents, technical reasoning, and conversational tasks without sacrificing raw text quality. The model is engineered for low-latency interactive use, with reported inference speeds around 150 tokens per second and an instruction-tuned design that targets chat, programming assistance, mathematical problem solving, and document comprehension. Its multimodal design lets it process both text and visual inputs together, while a multilingual training base spanning dozens of languages broadens its usefulness for global applications.

The release represents an evolution of the Mistral Small family rather than a wholesale retrain, refining the prior version with improved text performance, vision understanding, and longer-context behavior, and it is published as an instruction-finetuned variant of a public pretrained base. The accompanying material highlights strong results on reasoning benchmarks such as GPQA Diamond, knowledge suites including MMLU and MMLU-Pro, reading comprehension like TriviaQA, and quantitative and coding tasks such as MATH and HumanEval, where it is positioned as outperforming comparable small proprietary systems including Gemma 3 and GPT-4o Mini. Native function calling and structured JSON output give it agent-ready behavior suited for fast-response assistants and low-latency tool use, and its compact parameter count makes it practical for local inference on a single high-end consumer GPU or a quantized laptop setup. The Apache 2.0 licensing of the underlying weights, combined with strong baseline quality, makes it a flexible foundation for domain-specific fine-tuning, private on-device deployments, and enterprise workflows that need sensitive data to remain local.