Leanstral 1.5 119B A6B

Leanstral 1.5 119B A6B

An open code-agent model from Mistral for the Lean 4 proof assistant. It proves theorems and fixes Lean code, takes text and images, and runs through the Mistral API or your own vLLM server.

Made by mistralai on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Leanstral 1.5 is a code agent model with open weights, aimed at Lean 4. Lean 4 is a proof assistant that can state both advanced mathematics and specifications for software. It belongs to the Mistral Small 4 family and updates the earlier Leanstral release.

It is a mixture-of-experts model with 128 experts, 4 active per token, and 119 billion parameters of which 6.5 billion are active. The context length is 256k tokens. It accepts text and images and produces text.

Who it suits: Lean 4 users who want an AI agent that proves theorems or repairs code in their project, and developers building agents around Lean.

Industry | Who uses it | How they use it | Example
Mathematics | Lean users and mathematicians | Have an agent attempt proofs of stated theorems | Proving a lemma inside a Lean project
Software verification | Verification engineers | Write and repair formal specifications | Fixing a failing proof in a spec
AI agents | Agent developers | Pair the model with Lean tooling over MCP | Connecting the model to lean-lsp-mcp
Education | Students | Get help with Lean exercises | Asking the agent to complete a proof step

Anyone who does not work with Lean 4 can skip it, since the model is built for that proof assistant.

Where to use it:

• The Mistral Vibe command line tool, running inside your Lean project directory.
• The VS Code terminal, which the card suggests so vibe and your code sit side by side.
• Your own vLLM server, if you prefer not to call the Mistral API.
• The lean-lsp-mcp tool, which the card suggests for agents that work with Lean.

Use case | Who | What they get back
Theorem proving | Lean users | Proof attempts for a theorem you state
Code repair | Developers | Fixes to Lean code in your project
Long tasks | Researchers | Hours of autonomous work on one hard problem

How to set it up: You need a Mistral account setup or a local vLLM server.

• A Mistral account. The free plan is enough, with Labs models enabled in the privacy settings
• A Mistral API key
• The Mistral Vibe command line tool, installed as Mistral's own guide describes
• A Lean 4 project to work in

1. Sign in to your Mistral account and enable the Labs models option.
2. Create an API key.
3. Install the vibe command line tool.
4. Start vibe, which asks for your key. If it does not, run the setup command.
5. Inside vibe, enter the install command for Leanstral, then exit.

vibe --setup

Inside vibe, type these two lines in turn:

/leanstall
exit

How to use it: Open a terminal in your Lean project and start the Lean agent. Check that Leanstral shows at the top. Recommended settings are temperature 1.0 and a context of 200k tokens or less.

vibe --agent lean

To approve the model's changes automatically, add the yolo flag, and be careful with it:

vibe --agent lean --yolo

Tool | What it does
vibe --setup | Asks for your Mistral API key if setup did not run on its own
/leanstall | Installs Leanstral inside vibe
vibe --agent lean | Starts the Lean agent
--yolo | Approves model changes without asking
lean-lsp-mcp | A tool that lets agents talk to Lean

Example: In your Lean project folder, start the agent and ask it to prove a theorem you have written, or to fix a broken proof. It can keep working on a hard problem for hours, then reports what it changed.

For newcomers and experts: If you are new to Lean, start by reading a short proof, then ask the agent to explain or extend it. Keep the default approval prompts on so you see every change.

For experts, the reasoning effort setting has two values: none turns reasoning off and high turns it on for complex prompts. A local vLLM setup needs vllm 0.24.0 or newer and a custom agent file in the vibe agents folder. The card advises using the Mistral API if local serving gives weak results.

You open a Lean 4 project
Start vibe with the lean agent
Leanstral reads your files and goal
It works on a proof or a fix
You review the changes it makes

Good to know: The points below cover limits and costs.

• Reasoning: use high for complex prompts and none to skip reasoning.
• License: Apache 2.0.
• Cost to run: a Mistral API key is required for the hosted route, and the card says the free plan is enough. A local server needs a lot of GPU capacity.
• Data it reaches: your project files and prompts go to the Mistral API unless you run your own vLLM server. The agent config also gives it a shell tool.
• Not for: auto-approving changes without care, or work outside Lean 4.

On Hugging Face, 13 November 2026

• Library: vllm
• Downloads in the last 30 days: 425
• Likes: 234
• License: apache-2.0
• Based on: mistralai/Leanstral-2603
• Quantized versions: 15
• Fine-tunes: 2
• Tags: vllm

Links

• Model: https://huggingface.co/mistralai/Leanstral-1.5-119B-A6B
• Files: https://huggingface.co/mistralai/Leanstral-1.5-119B-A6B/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download mistralai/Leanstral-1.5-119B-A6B

Then follow the usage in its model card. Nothing in this kit runs by itself.