LLM & ChatActive

Mathstral 7B v0.1

A math and science model built on Mistral 7B for problems that take several steps of reasoning. Mistral AI reports 56.6% on the MATH benchmark for it.

ChatLLMRAGMistral ChatFp32

Model ID

Mathstral-7B-v0.1

Provider

mistralai

Updated

Oct 24, 2026, 14:57

Support Wiro on G2

wiro playground—mistralai/Mathstral-7B-v0.1

Updated Oct 24, 2026, 14:57

Overview

Mathstral 7B v0.1 is a math and science model that Mistral AI built on Mistral 7B and released in July 2024, as part of its collaboration with Project Numina. It is meant for problems that need several steps of logical reasoning. Mistral AI describes it as an instructed model, to be used or fine-tuned as one.

On MATH, Mistral AI reports 56.6% for this model, and 63.47% on MMLU. More work at answer time raises the MATH score: 68.37% with majority voting, and 74.59% when a strong reward model picks among 64 candidates.

Follow-up questions work too. With a user_id and session_id set, the earlier problem stays in the history, and a student can ask why one step holds.

What you can build

A tutoring app that walks a student through an equation one line at a time.
Worked solutions for practice sets, drafted by the model and checked by a teacher before release.
Homework helpers for physics and chemistry that rearrange formulas and convert units.
An answer checker that runs the same problem several times and keeps the answer that comes up most often, the approach behind Mistral AI's majority-voting score.

Inputs

The prompt is the only required field. Write the problem in plain text and use LaTeX where the notation needs it. The prefilled example, written with LaTeX, asks for the x that satisfies 4x + 5 = 6x + 7.

The other fields are optional. The system prompt starts with a general assistant instruction and is the natural place to fix how the final answer should look. Temperature (0.7) and seed (123456) come preset, next to the usual generation controls. This page has no function-calling fields.

Outputs

Text. The sample on this page solves 4x + 5 = 6x + 7 in three numbered steps (subtract 4x, subtract 7, divide by 2), writes out each intermediate equation and ends with x = -1.

Recommended settings

For a majority vote, keep sampling on, change the seed between runs and compare the final answers.
If code has to read the result, ask for the final answer on its own last line, for example one starting with "Answer:".
Mistral AI ships the weights in FP32. Quantization is on by default; switch it off to run the model at full precision.

Limitations

A 56.6% MATH score means many competition problems still come back wrong. Check answers before you mark them correct.
AIME-level problems are mostly out of reach: Mistral AI's table shows 2 of 30 AIME 2024 problems solved with majority voting over 16 samples.
The model cannot run code or use a calculator here, so every number in a solution is its own arithmetic.

Safety & compliance

Mistral AI releases Mathstral 7B v0.1 under the Apache 2.0 license.

API quick start

Run Mathstral-7B-v0.1 with a single API call.

POST https://api.wiro.ai/v1/Run/mistralai/mathstral-7b-v0-1

{
  "prompt": "Find the value of $x$ that satisfies the …",
  "user_id": "...",
  "session_id": "...",
  "system_prompt": "You are a helpful, respectful and honest …"
}

curl

curl -X POST "https://api.wiro.ai/v1/Run/mistralai/mathstral-7b-v0-1" \
  -H "Content-Type: application/json" \
  -H "x-api-key: YOUR_WIRO_API_KEY" \
  --data-binary @- <<'JSON'
{
  "prompt": "Find the value of $x$ that satisfies the …",
  "user_id": "...",
  "session_id": "...",
  "system_prompt": "You are a helpful, respectful and honest …"
}
JSON