Mistral Small 3.2
2025-06-20 · Publisher : Mistral AI · Domain : Artificial Intelligence
Mistral AI releases Mistral Small 3.2 (24B), a minor update to Small 3.1 under the Apache 2.0 license: better instruction following, fewer repetitions and more robust function calling.
What changes
- Better instruction following: Wildbench v2 from 55.6% to 65.33% and Arena Hard v2 from 19.56% to 43.1% according to the card.
- Infinite generation roughly halved on long, repetitive prompts (2.11% to 1.29%).
- More robust function calling template.
- Other capabilities equivalent to or slightly better than Small 3.1.
### Specifications - 24 billion parameters; 128k context; text and image input. - License: Apache 2.0 (weights on Hugging Face). - Availability: API since June 23, 2025, Hugging Face. - API identifier: `mistral-small-2506` (deprecated on 04/30/2026 according to the documentation). ### Pricing | Model | Input (USD / M tokens) | Output (USD / M tokens) | |---|---|---| | Mistral Small 3.2 | 0.1 | 0.3 | Prices from the pricing page archived on July 30, 2025. ### Benchmarks Scores published by the vendor (Hugging Face card, Small 3.2 versus Small 3.1): | Benchmark | Small 3.2 | Small 3.1 | |---|---|---| | MMLU-Pro (5-shot CoT) | 69.06% | 66.76% | | HumanEval Plus (pass@5) | 92.90% | 88.99% | | MBPP Plus (pass@5) | 78.33% | 74.63% | | SimpleQA | 12.10% | 10.43% | | ChartQA | 87.4% | 86.24% | ### Key takeaways A direct replacement for Small 3.1, more reliable for agents and tool calls, with no change in license or price.
Links
Tags : LLM, Mistral, Small, Open Source, Apache 2.0