Overview
Mathstral 7B is an open-weight large language model released by Mistral AI on 16 July 2024, fine-tuned from Mistral 7B and specialized for mathematical reasoning and STEM problem-solving. Named as a tribute to Archimedes, it was produced in collaboration with Project Numina and shipped as a community contribution: the weights are hosted on Hugging Face (mistralai/Mathstral-7B-v0.1) under the permissive Apache 2.0 license, runnable with mistral-inference and adaptable with mistral-finetune.
The model is a 7.3-billion-parameter instructed model with a 32K-token context window. Mistral positioned it as state-of-the-art for its size class on math benchmarks, scoring 56.6% on MATH and 63.47% on MMLU. Because it leaves more compute available at inference time, Mathstral also benefits from test-time scaling: with majority voting over 64 candidates it reaches 68.37% on MATH, and 74.59% with a strong reward model over the same candidates.
Mathstral was one half of a two-model 'specialist' launch alongside Codestral Mamba, a 7B code model built on the Mamba2 state-space architecture (API id codestral-mamba-2407, tested for in-context retrieval up to 256K tokens). Both were released the same day under Apache 2.0. Mathstral is now listed among Mistral's legacy/deprecated models, with Magistral Small suggested as a current alternative, but the open weights remain available for self-hosting and fine-tuning.
| Released | 2024-07-16 |
|---|---|
| License | Apache 2.0 |
| Weights | Open weights |
| Parameters | 7.3B |
| Context | 32K tokens |
| Max output | Not separately published (shares the 32K context budget) |
| Architecture | Dense transformer, fine-tuned from Mistral 7B. Mathstral keeps the Mistral 7B backbone (grouped-query attention, sliding-window-style efficient attention) and is specialized for mathematical and STEM reasoning. Its launch sibling, Codestral Mamba, instead uses a Mamba2 state-space architecture for linear-time inference. |
| Knowledge cutoff | Not published by Mistral |
| Modalities | text |
| Status | Deprecated. Mathstral 7B is listed in Mistral's legacy/deprecated models (recommended migration: Magistral Small). The open weights remain freely downloadable on Hugging Face under Apache 2.0. Its sibling, Codestral Mamba, was retired from the hosted API on 6 June 2025. |
Benchmarks
- MATH56.6%
- MMLU63.47%
- GSM8K (8-shot)77.1%
- MATH (maj@64, majority voting)68.37%
- MATH (reward model, 64 candidates)74.59%
- GRE Math (maj@16)56.9%
- AMC 2023 (maj@16)42.4%
- Odyssey Math (maj@16)37.2%
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Strengths
- Strong math reasoning for a 7B model — 56.6% on MATH, the best in its class at launch
- Fully open weights under Apache 2.0, free to self-host, modify, and fine-tune commercially
- Benefits from inference-time scaling (majority voting / reward model) for higher MATH accuracy
- Small enough to run on a single consumer GPU, making STEM tutoring and research experiments cheap
- Released as a fine-tunable base for academic and scientific projects (Project Numina collaboration)
Best for
- Solving multi-step math and STEM word problems
- Academic and research projects on mathematical reasoning
- Math tutoring and step-by-step explanation assistants
- Fine-tuning a small open model for domain-specific quantitative tasks
- Local/offline math reasoning where data cannot leave the device
How to access
| Provider | Model ID |
|---|---|
| Mistral AI (la Plateforme) ↗ | mathstral (legacy/deprecated) |
| Hugging Face (open weights) ↗ | mistralai/Mathstral-7B-v0.1 |
| Ollama ↗ | mathstral |
Specialist open models (Mathstral / Codestral Mamba) — every version
The full lineage of the Specialist open models (Mathstral / Codestral Mamba) line, newest first. Every version has its own page — click any to compare specs, benchmarks and pricing.
| Version | Released | Context | License |
|---|---|---|---|
| Mathstral 7Bcurrent | 2024-07-16 | — | Apache-2.0 |
| Codestral Mamba | 2024-07-16 | — | Open weights |
FAQ
Is Mathstral 7B free to use?
Yes. Mathstral 7B is released under the Apache 2.0 license, so you can download, run, modify, fine-tune, and use it commercially for free. The weights are on Hugging Face at mistralai/Mathstral-7B-v0.1, and it runs locally via mistral-inference or tools like Ollama. Note the model is now in Mistral's legacy/deprecated list.
What is Mathstral good at compared to regular Mistral 7B?
Mathstral is fine-tuned from Mistral 7B specifically for mathematical and STEM reasoning. At launch it scored 56.6% on MATH and 63.47% on MMLU, leading its 7B size class on math benchmarks. It also gains accuracy with inference-time techniques, reaching 68.37% on MATH with majority voting over 64 candidates.
How is Mathstral different from Codestral Mamba?
They were released the same day (16 July 2024) as a two-model specialist launch. Mathstral is a dense transformer focused on math, while Codestral Mamba is a code model built on the Mamba2 state-space architecture (tested for in-context retrieval up to 256K tokens). Both are 7B and Apache 2.0; Codestral Mamba was retired from the hosted API on 6 June 2025.
Is Mathstral still supported?
Mathstral appears in Mistral's legacy/deprecated models, with Magistral Small suggested as a current alternative for reasoning tasks. The open weights remain freely available on Hugging Face for self-hosting and fine-tuning, but it is no longer a primary supported model on la Plateforme.