Mistral Small 3.2 24B Instruct (2506)

Mistral Small 3.2 24B represents a targeted refinement of the earlier 3.1 release, built as a compact 24-billion parameter vision-language model designed for efficient multimodal reasoning. The model inherits the architecture that brought strong image understanding alongside text processing, but focuses on practical workflow improvements rather than wholesale changes. This position as a minor update means existing deployments can upgrade with confidence, gaining measurable gains without retraining or architectural migration.

The 3.2 iteration delivers meaningful advances in three practical areas: it follows complex, multi-step instructions more accurately, produces fewer repetitive or infinite generation loops on challenging prompts, and offers a more robust function-calling template for tool-use applications. Benchmark gains back these claims substantively—WildBench scores jump from 55.6% to 65.33% and Arena Hard rises from 19.56% to 43.1%, while reducing infinite generations by 2x on difficult, long, and repetitive inputs. The model performs well across coding tasks, STEM reasoning, and visual document understanding, making it versatile for production applications that require structured outputs and reliable tool orchestration alongside multimodal comprehension.