Key takeaways

• Mistral Small 3.2 is a 24B parameter, Transformer-based model optimized for instruction following, repetition reduction, and improved function calling, aiming to deliver gains in tool use and structured output tasks.
• The model demonstrates strong accuracy (97.77%) on 'easy' difficulty scientific reasoning problems (AI2 Reasoning Challenge - Easy dataset), showing a good understanding of scientific concepts and consistent performance across science domains.
• Significant improvements over its 3.1 predecessor include enhanced accuracy on WildBench and Arena Hard, reduced infinite generations, and gains in tool use and structured output tasks.
• Known limitations include struggles with complex reasoning, multi-step numerical calculations (e.g., in Biology/Medicine, 4.42% accuracy), integrating information from images, and often failing to adhere to specified output formats.
• Common failure modes include incorrect numerical calculations despite sound reasoning, incomplete or abruptly ending responses, inability to correctly integrate image data, and overconfidence in incorrect answers.

Technical specifications

Context length

96k tokens

Max output

—

Architecture

Transformer

Modalities

text->text

Parameters

~24B

Open weights

Yes

Licence

Apache 2.0

Regions

France

Vendor

MistralAI

Provider model key

mistralai/mistral-small-3.2-24b-instruct-2506-or