Mistral-Small-3.2-24B-Instruct-2506 is a visual language model that brings multimodal reasoning to a compact, accessible footprint. It extends its predecessor lineage by adding state-of-the-art vision understanding while keeping the 24-billion-parameter scale that balances strong capabilities with reasonable resource demands. The model supports inputs combining text and images, reasoning across visual content with reportedly top-tier performance in tasks like chart and document understanding. Long context handling reaches up to 128k tokens without apparent degradation in text performance, making it suitable for document-heavy workflows and multi-turn conversations that involve attached images or figures.

The 3.2 release is a focused refinement over Mistral Small 3.1, with demonstrable gains in instruction following accuracy, a more robust function calling template, and meaningfully reduced repetition or infinite generation issues. OpenRouter notes performance improvements on WildBench and Arena Hard benchmarks, alongside strong scores across coding tasks (HumanEval+, MBPP), STEM reasoning (MMLU, MATH, GPQA), and vision benchmarks (ChartQA, DocVQA). The model is distributed as open weights under Apache 2.0, inviting local deployment and community fine-tuning. Practitioners running it locally have found the 24B parameter size practical for laptop deployment, while its instruction finetuning lineage from the base model supports structured output, tool use, and temperature-controlled generation suitable for agentic applications.