Last updated 07-27-2026
Category:
Reviews:
Join thousands of AI enthusiasts in the World of AI!
Math??tral - Mistral AI
Mistral AI released Math??tral as an open-weight large language model focused on STEM reasoning and multi-step math problems. The 7B model builds on Mistral 7B and handles logical chains that trip up general chat models. Weights are on HuggingFace, and you run or fine-tune them locally with mistral-inference or mistral-finetune instead of paying per token through Mistral's API.
Where Mistral's flagship models chase broad capability, Math??tral bets on math and science depth in a small footprint. It scored 56.6% on MATH at default settings and climbs to 74.59% with majority voting plus a reward model selecting among 64 candidates. That test-time scaling pattern is unusual for a 7B release and reflects the academic math focus behind Mistral's Project Numina collaboration.
Researchers, ML engineers, and educators who need a compact model for benchmark work or custom math tutors get a free weight download with documented MMLU gains in STEM subjects over base Mistral 7B. Mistral developed it as a science community contribution, with GRE Math Subject Test problems curated by Professor Paul Bourdon for evaluation.
56.6% MATH benchmark at default inference, climbing to 74.59% with reward-model selection among 64 candidates
7B parameters built on Mistral 7B with documented STEM gains on MMLU over the base model
Open weights hosted on HuggingFace as mistralai/mathstral-7B-v0.1 for local download
Runs via mistral-inference v1.2.0+ or adapts with the mistral-finetune toolkit
Developed with Project Numina to tackle advanced problems needing multi-step logical reasoning
Free open weights on HuggingFace with no per-token API cost for self-hosting.
Strong MATH scores for a 7B model, with a clear path to higher accuracy via test-time voting.
Fine-tunable through mistral-finetune or la Plateforme for domain-specific math workflows.
Not offered as a hosted Mistral API model, so you must run or deploy the weights yourself.
STEM specialization means weaker general chat breadth than Mistral's larger flagship models.
Commercial production deployments of open-weight Mistral models may require a separate Mistral license.
Is Math??tral free to use?
Yes. Math??tral is a free open-weight model. Mistral hosts the mathstral-7B-v0.1 weights on HuggingFace for download, and you can run or fine-tune them locally without API token charges for the model itself.
How do you run Math??tral locally?
Math??tral runs through mistral-inference v1.2.0 or later after you download the HuggingFace weights. Mistral documents it as an instructed model, so follow their inference or mistral-finetune guides rather than treating it like a raw base checkpoint.
What MATH benchmark score does Math??tral hit?
Math??tral scores 56.6% on MATH at default inference. With majority voting it reaches 68.37%, and with a strong reward model picking among 64 candidates it hits 74.59% on the same benchmark.
Is Math??tral available on the Mistral API?
No. Math??tral is not listed on Mistral's API pricing page. You download the open weights from HuggingFace and self-host with mistral-inference or fine-tune with mistral-finetune instead of calling a hosted endpoint.
How big is the Math??tral model?
Math??tral is a 7B-parameter instructed model built on Mistral 7B. Mistral positions it in the small-model category where it reports leading reasoning scores for its size class on math benchmarks.
Who collaborated with Mistral on Math??tral?
Math??tral was produced in collaboration with Project Numina, an academic effort focused on advanced mathematical problem solving. Mistral also credits Professor Paul Bourdon for curating GRE Math Subject Test problems used in evaluation.
Can you fine-tune Math??tral?
Yes. Math??tral is an instructed model you can fine-tune with mistral-finetune or through fine-tuning capabilities on Mistral's la Plateforme. Mistral highlights purpose-built small models as a way to trade speed for domain depth.