Mistral Small 3.2 Lands With Sharper Function Calling and a 128K Context Window

Mistral has unveiled Mistral Small 3.2, a 24-billion parameter model that offers improved function calling capabilities and a 128K context window. This open-weight model competes with major players like OpenAI's GPT-4o mini and Anthropic's Claude Haiku, which have recently rolled out pricing changes and access restrictions.

Notable enhancements in version 3.2 include a more reliable tokenizer-side tool schema handling, enabling accurate respect for required vs optional distinctions and cleaner JSON output across multi-turn conversations. Additionally, the default sliding attention approach in vLLM 0.6.3+ boosts throughput for the Small line.

Function calling in Mistral Small 3.2 utilizes the v3 tool schema format, which aligns more closely with OpenAI's specification than the previous version. Running the model locally on a single H100 with vLLM, teams can expect around 80-100 tokens per second, with the marginal cost per token depending on electricity rates. For production environments, quantized builds are recommended.

Mistral Small 3.2 demonstrates competitive performance against GPT-4o mini on various reasoning benchmarks, such as MMLU-Pro, HumanEval, and IFEval, achieving scores within 3-5% of the smaller model. However, the real advantage lies in its function-calling reliability, which has improved from 71% in Small 3.1 to 78% in the latest version.

The open-weight nature of Mistral Small 3.2 sets it apart from closed models, as it allows for fine-tuning, distillation, and deployment without additional fees or usage attribution. This makes the model particularly cost-effective for startups handling 50 million tokens daily, potentially reducing costs by one-third compared to GPT-4o mini's API pricing. Furthermore, the availability of base models for other languages enables domain-specific fine-tuning, potentially recovering 90%+ of the original benchmark scores.