Summary:
1. Mistral, a French AI company, released an update to its 24B parameter open source model Mistral Small, focusing on improving instruction following and output stability.
2. The new version, Mistral Small 3.2, offers enhancements in behavior and reliability without major architectural changes, making it more efficient for developers.
3. Mistral Small 3.2 is available under the Apache 2.0 license, supported by frameworks like vLLM and Transformers, and provides a cleaner user experience for enterprises building on the Mistral ecosystem.
Article:
Mistral, a prominent AI company based in France, has been on a roll this summer with a series of new releases. Following the recent launch of its AI-optimized cloud service, Mistral Compute, the company has introduced an update to its 24B parameter open source model Mistral Small. The latest version, Mistral Small 3.2-24B Instruct-2506, aims to enhance specific behaviors such as instruction following, output stability, and function calling robustness. While the overall architecture remains unchanged, targeted refinements have been made to improve both internal evaluations and public benchmarks.
Mistral AI claims that Mistral Small 3.2 is better at adhering to precise instructions and reduces the likelihood of infinite or repetitive generations, which were occasional issues in previous versions when handling long or ambiguous prompts. The function calling template has also been upgraded to support more reliable tool-use scenarios, especially in frameworks like vLLM. Additionally, the new version can run on a single Nvidia A100/H100 80GB GPU, making it more accessible for businesses with limited compute resources or budgets.
The previous iteration, Mistral Small 3.1, was released in March 2025 as a flagship open release in the 24B parameter range. It offered full multimodal capabilities, multilingual understanding, and long-context processing of up to 128K tokens, positioning itself against proprietary competitors like GPT-4o Mini and Claude 3.5 Haiku. Small 3.1 emphasized efficient deployment, running inference at 150 tokens per second and supporting on-device use with 32 GB RAM.
In contrast, Mistral Small 3.2 focuses on fine-tuning behavior and reliability without introducing new capabilities or architectural changes. The update addresses edge cases in output generation, improves instruction compliance, and refines system prompt interactions. Instruction-following benchmarks demonstrate a measurable improvement, with Mistral’s internal accuracy rising from 82.75% in Small 3.1 to 84.78% in Small 3.2.
Performance on external datasets like Wildbench v2 and Arena Hard v2 has also seen significant improvements, with Mistral Small 3.2 outperforming its predecessor across various tasks. While some benchmarks show gains, others present a more nuanced picture, with minor fluctuations in vision benchmarks. Despite minor regressions in certain areas like the Massive Multitask Language Understanding benchmark, Mistral Small 3.2 remains a refinement focused on enhancing reliability and task handling.
Both Mistral Small 3.1 and 3.2 are available under the Apache 2.0 license and can be accessed via popular AI code sharing repository Hugging Face. The new version, Mistral Small 3.2, is supported by frameworks like vLLM and Transformers, requiring approximately 55 GB of GPU RAM to run in bf16 or fp16 precision. Developers can find system prompts and inference examples in the model repository for building or serving applications.
In conclusion, Mistral Small 3.2 may not revolutionize the open-weight model space, but it underscores Mistral AI’s commitment to iterative model refinement. With improvements in reliability and task handling, particularly around instruction precision and tool usage, Small 3.2 offers a smoother user experience for developers and enterprises within the Mistral ecosystem. Its compliance with EU regulations and French origin make it an attractive option for businesses in that region. While Mistral Small 3.1 remains a benchmark for performance in some areas, Mistral Small 3.2 serves as a stability-focused update with targeted enhancements for specific use cases.
