What is Leanstral 1.5?
Leanstral 1.5 is a mixture-of-experts model from the Mistral Small 4 family, tuned for Lean 4, the proof assistant. It reads text and images, writes text, and works across a 256K-token context. Mistral trained it to act like a developer inside a Lean project: it edits files, runs commands, reads compiler errors and keeps going, sometimes for millions of tokens, until the proof compiles.
It replaces the first Leanstral, which Mistral has retired. The hosted version sits in Mistral Labs, the experimental tier: free, but open to silent updates and removable at a month's notice. The Apache 2.0 weights stay yours.
What changed from the first Leanstral?
Mostly results on real proof work: the same size, further training on top of the first model, and better scores on Mistral's repository benchmark.
- Model
- Fine-tuned from the first Leanstral, with the same 119B total and 6.5B active parameters.
- Real repositories
- On FLTEval, built from real pull requests to a Lean project, pass@1 rose from 21.9 to 28.9 and pass@8 from 31.9 to 43.2.
- API
- The old labs-leanstral-2603 model is retired; new work goes to labs-leanstral-1-5.
Who is Leanstral 1.5 best for?
Leanstral 1.5 is best for Lean 4 users who want an agent to finish and repair proofs, formal maths researchers, engineers verifying Rust code, and teams that must self-host.
Lean 4 proof engineering
It finishes partial proofs, builds helper lemmas and fixes broken ones inside real repositories, the job Mistral trained it for.
Formal maths research
Mistral reports 100% on miniF2F, 587 of 672 PutnamBench problems and 87% on FATE-H, a graduate-level algebra set.
Verifying Rust code
Paired with Aeneas, which turns Rust into Lean, it flagged 11 real bugs across 57 open-source repositories in Mistral's test, 5 of them unreported.
Teams that must self-host
The Apache 2.0 weights run on your own GPUs with vLLM, so proofs and code never leave your servers.
What's included with Leanstral 1.5?
A Lean agent in Mistral Vibe, Lean language server tools, long runs across compactions, a reasoning switch, standard API features and open weights.
- Lean agent in Mistral Vibe
- Mistral's Vibe terminal agent has a Lean mode that installs with one command and works inside your project.
- Lean language server tools
- With the optional lean-lsp-mcp server it can read goals, errors and types while it works.
- Long runs
- It keeps working through context compactions; one AVL-tree proof ran for over 2.7 million tokens and 22 compactions.
- Reasoning switch
- Reasoning effort is either none, for quick answers, or high, for hard proofs.
- Standard API features
- Chat completions, function calling, agents and structured outputs all work on the Labs endpoint.
- Open weights
- The 119B weights are on Hugging Face and need about 238 GB of GPU memory at bf16, or about 60 GB at fp4.
How does Leanstral 1.5 work?
Get a key
On Mistral's free plan, turn on Labs models in the admin privacy settings and create an API key.
Install Vibe
Install Mistral Vibe, enter your key, then run /leanstall to add Leanstral.
Start the Lean agent
Run vibe --agent lean inside your Lean project, ideally with the lean-lsp-mcp server added.
Hand it a proof
Ask it to prove a theorem, debug a proof or fill gaps in a repository, and let it run.
How much does Leanstral 1.5 cost?
Leanstral 1.5 is free: Mistral's API charges nothing for it for a limited period, and the Apache 2.0 weights cost nothing to download and run on your own hardware.
- Model ID labs-leanstral-1-5
- Free for a limited period; works on Mistral's free plan once Labs models are on
- No opt-out from data collection on Labs models
- 119B parameters, 6.5B active per token
- About 238 GB of GPU memory at bf16, 60 GB at fp4
- Served with vLLM
Mistral says it keeps the endpoint free for a limited period to gather feedback and usage data, and Labs models can be changed or removed with one month's notice. Mistral has not published a paid price for Leanstral. Full pricing on mistral.ai. We checked these prices on .
How do readers rate Leanstral 1.5?
No one has rated Leanstral 1.5 yet. Be the first.
Sign in to rate. One rating per account.
Leanstral 1.5 pros and cons
It is free, open and strong on Mistral's proof benchmarks for its small active size, but the hosted model is an experimental Labs release and general frontier agents can still beat it.
Pros
- Free on Mistral's API for now, and the Apache 2.0 weights can be downloaded and run with no licence fee.
- Beats every prover in Mistral's chart except Aleph Prover on PutnamBench, with 587 of 672 solved, 7 more than Seed-Prover 1.5 high at about $4 a problem, and it tops the chart on FATE-H and FATE-X.
- Only 6.5B of its 119B parameters are active per token, which Hacker News commenters saw as the real news: proof work at this level from a model small enough to host yourself.
- Scales with budget: Mistral shows PutnamBench solves rising from 244 at 200K tokens per attempt to 587 at 4M.
Cons
- The hosted model is a Labs release: free only for a limited period, open to silent updates, removable at a month's notice, not meant for production, and with no opt-out from data collection.
- Not the strongest prover overall: Aleph Prover solves more PutnamBench problems in Mistral's own chart, and Hacker News users note it trails some rivals, with one prover-tool author finding Claude Code and Codex still ahead in his tests.
- The free API needs an extra step, turning on Labs models in Mistral's admin settings, and several early users hit errors there until Mistral fixed it.
- It only works in Lean 4, and self-hosting is not a laptop job: even at fp4 the weights need about 60 GB of GPU memory.
Is Leanstral 1.5 open source?
Yes. Leanstral 1.5's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Leanstral 1.5 alternatives?
Leanstral 1.5 is an open source alternative to Gauss.
GaussPaidAn autoformalization agent from Math, Inc. that turns mathematics into formal Lean proofs.
Nomos 1Open SourceChoose it when you want maths proofs written in plain language rather than in Lean, from a smaller open model: a 30B Apache 2.0 model built on Qwen3 that runs with Nous Research's own reasoning harness.
Claude Opus 5.5PaidChoose it when the proofs sit inside a larger software project and you want one production model for the Lean and the code around it; it costs $4 / $20 per million tokens and is not an experimental release.
Gemini 3.1 Deep ThinkPaidChoose it when you want to reason through research-level maths in plain language in the Gemini app rather than get machine-checked proofs; it comes with Google AI Ultra, from $99.99 a month.
Questions about Leanstral 1.5
What is Leanstral 1.5?
Leanstral 1.5 is Mistral AI's open-weight model for mathematicians and engineers who work in Lean 4; it writes and repairs formal proofs as an agent, and its API is free for now.
Who is Leanstral 1.5 for?
Leanstral 1.5 is best for Lean 4 users who want an agent to finish and repair proofs, formal maths researchers, engineers verifying Rust code, and teams that must self-host.
Is Leanstral 1.5 free?
Yes, for now. Mistral charges nothing for labs-leanstral-1-5 on its API for a limited period, and the Apache 2.0 weights are free to download and run. Mistral has not said what the hosted model will cost when the free period ends.
Is Leanstral 1.5 open source?
Yes. Leanstral 1.5's code is public under the Apache 2.0 licence, so you can run it on your own server.
What are the best Leanstral 1.5 alternatives?
The closest alternatives on Recaply are Gauss, Nomos 1, Claude Opus 5.5 and Gemini 3.1 Deep Think.
Where can I try Leanstral 1.5?
On its official website, mistral.ai.
Is Leanstral 1.5 ready for production use?
Not on Mistral's API. Mistral says Labs models are experimental, can be updated silently and can be removed with a month's notice, and it does not recommend them for production. If you need a fixed version, run the Apache 2.0 weights yourself.
Does Mistral use the code I send to Leanstral 1.5?
It can. Mistral's lifecycle policy says Labs models do not support opting out of data collection, and the pricing page says the free endpoint exists partly to gather feedback and usage data. Keep private code on a self-hosted copy of the weights.
Can I run Leanstral 1.5 on my own computer?
Only with serious hardware. Mistral estimates about 238 GB of GPU memory at bf16 and about 60 GB at fp4, and its vLLM example spreads the model across four GPUs. It is not a laptop model.
Do I need a paid Mistral plan to use Leanstral 1.5?
No. Mistral's model card says the free plan is enough: turn on Labs models in the admin console's privacy settings, create an API key and connect it to Mistral Vibe or your own client.
Our verdict
Leanstral 1.5 is the cheapest serious way to put a Lean 4 proof agent to work: it costs nothing.
It posts strong scores on Mistral's proof benchmarks from a model with only 6.5B active parameters, and both the API and the Apache 2.0 weights are free.
- Skip it for
- Work that needs a stable, private hosted service, since the API is an experimental Labs release with no data opt-out.
- Check first
- If privacy or uptime matter, check that you can host the weights, which need about 60 GB of GPU memory even at fp4.
Hunyuan Vision 1.5 Thinking
Brave Ocelot
Hydrogen-1.0
Biohub ESM
MiroThinker
Interfaze 1.1 Lite