Mistral Small 3.2 24B Instruct is a 24-billion-parameter instruction-finetuned model built as a direct refinement of the Mistral Small 3.1 instruction series. It brings state-of-the-art multimodal vision capabilities alongside strong text performance, and extends the context window up to 128,000 tokens for longer document processing. The model is designed with particular attention to instruction adherence, repetition control, and tool-use reliability. Its function-calling template was hardened specifically to handle more robust tool interactions, and it supports structured outputs in both text and vision modes. With access to coding benchmarks like HumanEval Plus and STEM benchmarks including MMLU, MATH, and GPQA, the model targets developers and researchers who need reliable instruction-following, function calling, and visual document understanding at a scale that remains feasible to run on developer hardware.
Training lineage traces back through Mistral Small 3.1, which itself was an instruction-finetuned derivative of the base model Mistral Small 3.1 24B Base. The 3.2 update focuses on measurable quality gains: improved accuracy on WildBench and Arena Hard, fewer infinite generation loops, and stronger structured output behavior. Benchmark rankings show strong positioning in legal, finance, and document-vision tasks, with DocVQA and AI2D scores placing it near the top of vision leaderboards. The model is released as an open-weight package under the Apache 2.0 license, making it accessible for local deployment on developer machines. Mistral's recommended temperature of 0.15 reflects its emphasis on precise, controlled outputs rather than open-ended generation, giving users a tuned starting point for tasks like tool orchestration, structured data extraction, and document-based question answering.
