Mistral AI Releases MathΣtral, New Model for Math Reasoning and Scientific Discovery
Mistral AI has unveiled MathΣtral, a specialised 7B model designed for advanced mathematical reasoning and scientific exploration. Released under the Apache 2.0 license, MathΣtral pays homage to Archimedes on the occasion of his 2311th anniversary this year.
MathΣtral is tailored to tackle complex, multi-step logical reasoning challenges in STEM fields. Developed in collaboration with Project Numina, the model inherits capabilities from Mistral 7B, achieving state-of-the-art performance across industry-standard benchmarks. Notably, it scores 56.6% on MATH and 63.47% on MMLU, demonstrating superior reasoning capacities within its size category.
Detailed benchmarks highlight MathΣtral’s robust performance improvements with increased inference-time computation. For instance, MathΣtral 7B achieves significant accuracy enhancements, scoring 68.37% on MATH through majority voting and 74.59% with a strong reward model among 64 candidates.
MathΣtral is available for immediate use and adaptation using Mistral AI’s tools. Developers can deploy the model through mistral-inference for initial exploration and fine-tune its capabilities with mistral-finetune. The model’s weights are accessible via HuggingFace, facilitating straightforward integration into academic and research projects.
By releasing MathΣtral to the scientific community, Mistral AI aims to foster advancements in mathematical problem-solving and support academic endeavors. This initiative underscores Mistral AI’s commitment to promoting specialized model architectures and their practical applications in scientific discovery.
