Leanstral-1.5-119B-A6B

Model release Hugging Face 模型 · mistralai Source language: Chinese

官方机构模型仓库创建于 2026-07-02。仓库创建时间不等于正式发布日期。任务类型:未提供

Full text

Leanstral 1.5 119B A6B

Leanstral 1.5 is an open-source code agent model designed for Lean 4, a proof assistant capable of expressing complex mathematical objects such as perfectoid spaces and software specifications like properties of Rust fragments.

Built as part of the Mistral Small 4 family, it combines multimodal capabilities and an efficient architecture, making it both performant and cost-effective compared to existing closed-source alternatives.

[!Note] This is an updated model of our previously released Leanstral model.

Key Features

Leanstral incorporates the following architectural choices:

MoE: 128 experts, 4 active per token Model Size: 119B parameters with 6.5B activated per token Context Length: 256k tokens Multimodal Input: Accepts text and image input, producing text output

Recommended Settings

Temperature: 1.0 Reasoning Effort: 'none' → Do not use reasoning 'high' → Use reasoning (recommended for complex prompts). Context Length: ≤ 200k tokens recommended

Usage

Installing Mistral Vibe and Leanstral

If you have already installed Leanstral it should automatically use the new model. Verify your Mistral account settings at https://chat.mistral.ai/ - Sign up or Sign in - You can stay on the free plan (there is no need to sign up to Mistral Pro to use Leanstral). - Opt into “Enable Labs models” in https://admin.mistral.ai/plateforme/privacy - Make an API key in https://admin.mistral.ai/organization/api-keys Follow the instructions to install the mistral vibe CLI at https://docs.mistral.ai/getting-started/quickstarts/vibe-code/install-cli Setup should happen automatically, requesting your API key. But if not, use vibe --setup from the terminal. To setup Leanstral in Mistral vibe enter: - /leanstall (lean + [in]stall) - exit (to leave vibe)

Using Leanstral

Type vibe --agent lean in the terminal (ensure Leanstral displays at the top). - We recommend running vibe inside vs code terminal so you can see both vibe and your code. - We recommend running vibe inside your Lean project directory. - To auto-approve model changes (be careful!), use vibe --agent lean --yolo. Ask Leanstral to do any coding task for you, such as prove a given theorem or fix some code in your project. Leanstral is capable of doing very long range tasks which requires hours of work. Don’t hesitate to let it work away at a problem.

Using Leanstral with the lean-lsp-mcp

We recommend trying Leanstral with the lean-lsp-mcp, a standard tool for AI agents to interact with Lean. The lean-lsp-mcp README has instructions for setting up the MCP with Mistral Vibe.

Local server

If instead of pinging the Mistral API, you want to use your local vLLM server, you can do the following: 1. Spin up a vllm server as explained in Usage - vllm 2. Create a new agent file called lean.toml in ~/.vibe/agents:

And then copy-paste the following config into ~/.vibe/agents/lean.toml

Note: Make sure to overwrite with your server's url.

Then restart vibe and "tab-shift" to "lean" mode.

Give it a try on some "lean" code such as, *e.g.*: PrimeNumberTheoremAnd

Local Deployment

The model can also be deployed with vLLM, we advise everyone to use the Mistral AI API if the model is subpar with local serving: vllm (recommended): See here.

vLLM

Let's use this model with the vLLM library to implement production-ready inference pipelines.

Installation

Make sure to install vllm >= 0.24.0:

Doing so should automatically install mistral_common >= 1.11.5.

To check:

You can also make use of a ready-to-go docker image or on the docker hub.

Launch server

We recommend that you use Leanstral in a server/client setting.

Client

_Example Content_:

Expand</summary To define transaction rules as an inductive proposition in Lean, you can create an inductive type that captures the valid transitions between states. This approach is particularly useful for modeling state machines and proving properties about transitions, such as invariants.

Here is an example of how to define an inductive proposition for valid transitions in a simple counter machine:

In this example, ValidTransition is defined as an inductive proposition with constructors step1, step2, and zero_step. The step1 constructor allows for any state to transition to the next state by incrementing by 1, while step2 allows for doubling the state. The zero_step constructor is a special case for the initial state 0.

The lemma reach_any_positive aims to prove that any positive natural number is reachable from the initial state 0 through valid transitions. The proof uses induction on n, but due to simplification issues, the current approach may not be the most effective.

A more concise and effective way to define and prove properties about transitions is to use a simpler model, such as the following:

In this simpler model, CounterTrans is defined with a single constructor valid, which allows for any state to transition to the next state by incrementing by 1. The lemma reach_positive is straightforward to prove by induction, leveraging the simplicity of the transition rules.

This approach ensures that the proof remains concise and effective, avoiding unnecessary complexity in the transition rules. By using inductive propositions, we can effectively reason about state transitions and prove properties about the system.

_Tool-Calling_

You can add tools to the chat completion as follows:

lean\ninductive State where\n idle\n busy\n error\n\ndef transition : State → State → Bool\n .idle, .busy => true\n .busy, .idle => true\n .busy, .error => true\n _, _ => false\n\n#eval transition .idle .busy\n

_Example Tool Calls_:

Expand true\\n .busy, .idle => true\\n .busy, .error => true\\n _, _ => false\\n\\n#eval transition .idle .busy"}', name='lean_run_code')

License

This model is licensed under the Apache 2.0 License.

*You must not use this model in a manner that infringes, misappropriates, or otherwise violates any third party’s rights, including intellectual property rights.*

Sources

← Back to the daily brief