Mistral-Small-3.2-24B-Instruct-2506

Overview. Mistral-Small-3.2-24B-Instruct-2506 is a 24B dense open-weights instruct model from Mistral AI, released in June 2025 as a minor update over Mistral Small 3.1. It accepts text and image input and returns text, ships under Apache-2.0, and improves instruction following and native function calling versus the prior checkpoint. Mistral now marks this release deprecated and directs new integrations to Mistral Small 4, but the weights remain publicly available for self-host and third-party Inference.

Key strengths. On published benchmarks the model posts MMLU 80.50%, MMLU-Pro 69.06%, MATH 69.42%, GPQA Diamond around 46–50.5%, and HumanEval Plus pass@5 92.90%, with Wildbench v2 at 65.33%. Vision scores are strong for a mid-size checkpoint: DocVQA 94.86%, ChartQA 87.4%, MMMU 62.5%. It supports native tool calling including parallel tool calls, structured output and JSON mode, a 131k context window, and multilingual coverage across two dozen languages.

When to use it. Reach for this model on mid-tier interactive chat, coding assistance, document and chart reading, and function-calling agents where medium complexity is enough and you want text-plus-image input in one pass. The interactive latency class (roughly 38 tok/s median, sub-second time-to-first-token on reported runs) fits multi-turn product surfaces. Domain fit is strongest on general, communication, writing, and code workloads; it is a weaker choice for sustained agentic loops.

Limitations. Mistral has deprecated the model in favour of Mistral Small 4, so it should not be the default for new long-lived integrations. Knowledge is cut off at 2023-10-01, there is no extended-thinking or reasoning mode, and agentic benchmark scores are low. Output is text only; audio input, native web search, code execution, and file search are not built in.

Sovereignty note. Built by Mistral AI in France with no CLOUD Act exposure, released as open weights under Apache-2.0, and available on EU-resident hosts as well as self-hosted. That combination qualifies as Sovereign for workloads that keep Inference inside EU-owned infrastructure.