AI BriefingMathstral
Key point
Mistral AI has unveiled **Mathstral**, a STEM-specialized model designed for complex mathematical reasoning.
Details
Mistral AI, in collaboration with Project Numina, has unveiled Mathstral, which supports complex multi-step logical reasoning in mathematics and STEM fields. Built on Mistral 7B, this model demonstrates top-tier reasoning capability in its class.
Key benchmark performance is as follows:
- MATH: 56.6%
- MMLU: 63.47%
In particular, performance improves significantly when inference-time computation is increased. On the MATH benchmark, applying majority voting achieved 68.37%, and using a strong reward model among 64 candidates reached 74.59%.
Mathstral is provided as an instructed model, and its weights are available on HuggingFace. Users can test the model via mistral-inference or fine-tune it with mistral-finetune.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.