Mathstral

What it is

Mathstral is a language model with 7 billion parameters, released by Mistral AI on 16 July 2024 and built for mathematical reasoning^[1]. Mistral described it as a model specialising in STEM subjects that stands on the shoulders of Mistral 7B, the company's earlier general-purpose model it is based on^[1]. In other words, it is not a large model trained from scratch but a small base model adapted to a narrower domain.

The published checkpoint stores its 7 billion parameters as F32 tensors^[3].

How it was made

Mistral AI developed Mathstral in collaboration with Project Numina^[1]. That collaboration is the only detail Mistral gave about the model's construction; the emphasis in the announcement is on what the model can do rather than on how it was trained.

Benchmark results

On MATH, a benchmark of competition-style mathematics problems, Mathstral scored 56.6%, and on MMLU, which tests general knowledge across a wide range of school and professional subjects, it scored 63.47%^[1].

Mistral also reported scores obtained by spending extra computation at inference time rather than training time. Majority voting — sampling several solutions to the same problem and keeping the answer that appears most often — raised the MATH score to 68.37%. Reranking 64 candidate solutions with a strong reward model raised it further, to 74.59%^[1]. The pattern is a familiar one: a small model's single-sample accuracy can be improved substantially by generating many attempts and then selecting among them.

The model card published alongside the weights reports additional figures. On GSM8K, a set of grade-school arithmetic word problems, Mathstral scored 77.1% in the 8-shot setting, where the model is shown eight worked examples before answering^[3]. With maj@16 — sixteen samples per question, with the majority answer taken as the result — it reached 37.2 on Odyssey Math, 56.9 on GRE Math, 42.4 on AMC 2023, and 2 out of 30 on AIME 2024^[3].

The spread between those numbers is itself the interesting part. A 7B model can handle school-level arithmetic and standard exam-style questions competently, while olympiad-level problems remain largely out of reach: two correct answers out of thirty is far from a solved benchmark.

Availability and licensing

The weights are hosted on Hugging Face^[1], and the model card lists the licence as apache-2.0^[3], which permits reuse and modification with few restrictions. Mathstral is now listed among Mistral's deprecated models, with Mistral Small 4 given as the suggested replacement^[2]. Deprecation in this sense means the company no longer presents it as a model to build new work on, not that the published weights have been withdrawn.

Where it sits

Mathstral belongs to a class of small, openly licensed domain models that labs released alongside their general-purpose flagship models. Its design assumption is that a compact base model plus targeted specialisation can be a practical tool for a specific kind of work — here, mathematics and related technical subjects — at a size that is cheap to run.

Its results also show the ceiling of that approach at this scale. Inference-time techniques such as majority voting and reward-model reranking moved MATH from 56.6% to 74.59%, a large gain obtained without retraining the model^[1], but the AIME 2024 figure of 2 out of 30 shows where the limit lies^[3]. For readers following the field, Mathstral is a useful reference point for how much a 7B model could and could not do on mathematical problems, and for how much of a reported score can come from how the model is queried rather than from the model itself.