Mistral Small 3.2 24B Instruct is positioned by Mistral AI as a minor but targeted update over the earlier 3.1 release, concentrating on three pain points reported by users of the previous version. The official model card highlights better adherence to precise instructions, a meaningful drop in infinite-generation and repetition failures on long or tricky prompts, and a more robust function-calling template, while leaving most other behavior largely unchanged. The OpenRouter listing for the same model reiterates that theme, describing the update as optimized for instruction following, repetition reduction, and improved tool use, with structured outputs available alongside those capabilities.
Beyond instruction behavior, the model is designed as a multimodal assistant: it accepts both text and image inputs and returns text, making it suitable for document understanding and vision-grounded tasks in addition to standard chat and reasoning work. Benchmark framing from the model card and the OpenRouter summary points to broad coverage across coding evaluations such as HumanEval+ and MBPP, STEM reasoning suites including MMLU, MATH, and GPQA, and vision benchmarks like ChartQA and DocVQA, with notable score gains on WildBench v2 and Arena Hard v2 over the prior 3.1 version. Practically, this profile makes Small 3.2 24B a sensible mid-sized choice for teams that need reliable tool use and structured output from a model that can also read images, without committing to a larger flagship deployment.
