Mathstral 7B Review: Mistral AI's Open Math LLM — Benchmarks & Access
Published July 16, 2024 · Updated June 15, 2026
TL;DR
Mathstral 7B (July 2024) is Mistral AI's open-weight math specialist, built on Mistral 7B and fine-tuned in collaboration with Project Numina. It achieves 56.6% on the MATH benchmark — state-of-the-art for 7B models at release — and scales to 74.59% with majority voting and a reward model. Weights are free on Hugging Face. Best suited for academic research and math-focused fine-tuning; for production, newer models have raised the bar significantly.
What is Mathstral 7B?
Mathstral 7B is an open-weight language model released by Mistral AI on July 16, 2024, purpose-built for advanced mathematical and STEM reasoning. Unlike general-purpose LLMs, Mathstral was designed specifically to handle complex, multi-step problems of the type found in math competitions, graduate coursework, and academic research.
The model is built on the Mistral 7B architecture — Mistral AI describes it as standing “on the shoulders of Mistral 7B” — and was produced in collaboration with Project Numina, an academic initiative focused on AI for mathematical problem solving. Mistral contributed Mathstral to the science community as a research artifact rather than a commercial offering.
The release was notable because it demonstrated that a 7B-parameter model, when fine-tuned with the right mathematical data and objective, could reach MATH benchmark scores previously associated with much larger models — and that inference-time scaling (majority voting + reward models) could push those scores substantially further.
Model Specifications
| Spec | Value | Note |
|---|---|---|
| Parameters | 7B | Built on Mistral 7B base architecture |
| Type | Instructed (SFT) | Fine-tuned for step-by-step math reasoning |
| MATH benchmark | 56.6% | State-of-the-art in the 7B class at release |
| MATH (scaled inference) | 74.59% | Majority voting + strong reward model among 64 candidates |
| MMLU overall | 63.47% | Strong gains on STEM subjects vs. base Mistral 7B |
| License | Research (Mistral Research License) | Free for academic and research use; check for commercial restrictions |
| Weights | Hugging Face | mistralai/mathstral-7B-v0.1 |
| Collaboration | Project Numina | Academic partnership for advanced mathematical reasoning |
Benchmark Performance
Mathstral's headline result is 56.6% on the MATH benchmark — a curated dataset of competition mathematics problems requiring multi-step logical reasoning. This was a meaningful improvement over the Mistral 7B base model and competitive with much larger models at the time of release.
More striking is the inference-time scaling result: 74.59% on MATH when using majority voting among 64 generated candidates scored by a strong reward model. Majority voting alone (without the reward model) yields 68.37%. This demonstrates that Mathstral is well-suited for applications that can afford multiple inference passes — competition math solvers, tutoring systems with verification, or research pipelines.
| Model | MATH | MATH (scaled) | MMLU | Note |
|---|---|---|---|---|
| Mathstral 7B | 56.6% | 74.59% | 63.47% | Mistral AI, July 2024 |
| Mistral 7B (base) | ~11% | — | 60.1% | Baseline; Mathstral is a fine-tune of this |
| Llama 3 8B | ~30% | — | 66.6% | Meta, April 2024 |
Benchmarks from Mistral AI's release post (July 2024). Mathstral MATH-scaled uses majority voting + strong reward model among 64 candidates.
What Makes Mathstral Different
Math-Specialized Fine-Tuning
Mathstral is an instructed model fine-tuned on mathematical data curated in collaboration with Project Numina. Rather than being a general chat assistant, it is optimized for generating step-by-step solutions to competition-level math problems — the kind of structured, verifiable reasoning needed for proofs and multi-step calculations. The GRE Math Subject Test problems used in evaluation were curated by Professor Paul Bourdon of the University of Virginia.
Inference-Time Scaling
One of Mathstral's most interesting properties is how well it benefits from inference-time compute scaling. By sampling 64 candidate solutions and using a reward model to pick the best, accuracy on MATH jumps from 56.6% to 74.59% — an 18-point gain. This makes Mathstral particularly well-suited for pipelines that can afford multiple inference passes and need correctness rather than speed.
Mistral 7B Compatibility
Because Mathstral is built on Mistral 7B, it is compatible with the existing mistral-inference and mistral-finetune toolchains. Teams already familiar with Mistral 7B deployment can run Mathstral with minimal infrastructure changes — simply swap the weights and use the instruct prompt format.
How to Access Mathstral
Hugging Face (Free Download)
Weights are available at mistralai/mathstral-7B-v0.1 on Hugging Face. Download and run locally using Hugging Face Transformers or the official mistral-inference library.
mistral-inference (Official CLI)
Mistral AI's mistral-inference library (v1.2.0+) supports Mathstral directly. Use the instruct format — not bare completion — for best results on math problems.
Fine-Tuning
Mathstral can be fine-tuned further using mistral-finetune. This is useful for adapting to specific math subdomains (e.g., physics, statistics, theorem proving) or proprietary problem sets.
Who Should Use Mathstral in 2026?
By mid-2026, Mathstral has been outpaced by several more capable math models at both the 7B and larger scales. For most new production projects requiring math reasoning, DeepSeek-Math, Qwen-Math, or frontier APIs (GPT-4o, Gemini 2.5, Claude) are stronger choices.
Mathstral still makes sense for:
- Academic researchers: Studying 2024-era math LLM capabilities or building on the Project Numina methodology.
- Inference-time scaling research: Mathstral's published numbers make it a useful baseline for studying reward model + majority voting approaches.
- Compute-constrained math applications: A self-hosted 7B model with solid MATH performance has zero API cost, useful for high-volume math tutoring systems on a budget.
- Mistral 7B migration paths: Teams already running Mistral 7B who need better STEM performance without switching architecture.
Frequently Asked Questions
What is Mathstral 7B?
Mathstral 7B is an open-weight language model released by Mistral AI on July 16, 2024, specialized for mathematical and STEM reasoning. It is built on the Mistral 7B architecture and fine-tuned with data and methodology developed in collaboration with Project Numina, an academic initiative focused on advanced mathematical problem solving. Mathstral was contributed to the science community as a research artifact rather than a commercial product.
What benchmarks does Mathstral achieve?
At release, Mathstral 7B scored 56.6% on the MATH benchmark (a competition math dataset requiring multi-step reasoning) and 63.47% on MMLU overall, with strong per-subject gains on STEM topics compared to the Mistral 7B base model. With inference-time scaling — majority voting among 64 candidates combined with a strong reward model — Mathstral reaches 68.37% on MATH with majority voting alone and 74.59% with the reward model. These were state-of-the-art results in the 7B parameter class at release in July 2024.
How does Mathstral compare to DeepSeek-Math or other math LLMs?
At release in mid-2024, Mathstral was competitive with the best open math models in the 7B class. However, by 2025–2026, larger and more capable math-specialized models — including DeepSeek-Math 7B (which scored similarly at MATH at release), Qwen-Math, and later Gemini 2.5 Flash — have significantly raised the bar. For new applications requiring high math accuracy, those models or frontier APIs (GPT-4o, Claude, Gemini) are generally stronger choices. Mathstral remains useful as a compact, self-hostable baseline.
Can I use Mathstral commercially?
Mathstral was released under the Mistral Research License, which is designed for academic and research use. It is more restrictive than the Apache 2.0 license used for Mistral NeMo and Mistral Small. Before using Mathstral in a commercial product, review the license terms on the Hugging Face model page (mistralai/mathstral-7B-v0.1) or contact Mistral AI for clarification. For commercial math applications, a Mistral API or a model under Apache 2.0 may be a better fit.
How do I run Mathstral?
Mathstral weights are available on Hugging Face at mistralai/mathstral-7B-v0.1. You can run inference using mistral-inference (Mistral AI's official library, available on GitHub at mistralai/mistral-inference) or with standard Hugging Face Transformers. Fine-tuning is supported via mistral-finetune. Because Mathstral is an instructed model, use the instruct prompt format (not bare completion) for math problems to get the best results.
What is Project Numina?
Project Numina (projectnumina.ai) is an academic initiative focused on using AI to advance mathematical reasoning, particularly competition mathematics and theorem proving. Mistral AI collaborated with Project Numina to produce Mathstral, contributing data curation and evaluation methodology. The GRE Math Subject Test problems used in Mathstral's evaluation were curated by Professor Paul Bourdon of the University of Virginia.
Is Mathstral still relevant in 2026?
For most production applications, no — newer models have substantially higher MATH benchmark scores. However, Mathstral retains niche value for: (1) researchers studying math LLM fine-tuning methodology from 2024, (2) teams that need a 7B-scale math-capable model for extremely cost-constrained deployment, (3) academic projects requiring an open baseline tied to a specific era, and (4) fine-tuning experiments where Mistral 7B architecture compatibility matters.
Try Mathstral 7B
Open-weight math LLM from Mistral AI — free to download and self-host from Hugging Face.
View Mathstral on Hugging FaceAffiliate disclosure: Some links on this page are affiliate links. If you sign up through them, AISO Tools may earn a commission at no extra cost to you. This never affects our rankings or reviews.
📬 Get the best new AI tools delivered weekly
One concise email with fresh launches, trending picks, and featured standouts.
Join thousands of professionals who discover the best AI tools every week. No spam — unsubscribe anytime.