On June 30, Mistral shipped Leanstral 1.5, an open-weight Apache 2.0 model built for formal proof work in Lean 4. We ran it for a week on three concrete use cases. Verdict: plug it in if you live in formal verification, ignore it entirely otherwise.
Key Takeaways
- Leanstral 1.5 is a 119B / 6.5B active MoE, 256k context, Apache 2.0, free in Mistral Labs
- It solves 587 of 672 PutnamBench problems, a serious formal-math benchmark
- Useful for labs, security auditors and math teachers; overkill for most tech teams
What Actually Changes With Leanstral 1.5
Leanstral 1.5 is the direct successor to the first Leanstral, Mistral’s dedicated model for Lean 4, a formal verification language used by mathematicians and a segment of security engineers. Two shifts matter: size and licensing.
On size, Mistral moves to a Mixture of Experts architecture. The model has 119 billion parameters total, 6.5 billion active per forward pass. The tradeoff is classic: more capacity, contained latency, controlled inference cost.
Context length jumps to 256,000 tokens. That is a bare minimum for handling proofs alongside their dependencies, lemmas and related code. This single change unlocks use cases the previous version could not handle cleanly.
On licensing, Mistral keeps the Apache 2.0 stance. Weights are on Hugging Face, commercial use is allowed, and the model is also free inside Mistral Labs for teams without the GPU footprint to run it locally. The move clearly targets the research community and embedded verification.
The strategy overlaps with other labs going vertical. Anthropic is opening Claude Science for research labs with a research targeting. Mistral is playing the same card, on a narrower slice: math and software verification.
Our Test on Three Real-World Cases
We ran Leanstral 1.5 across three scenarios that engineering teams actually hit, short of doing pure math research. The goal was to check where the model earns its place outside academia.
First case: formalizing a simplified smart contract (conditional transfer, no reentrancy, invariant on the total supply). On this class of object, Leanstral 1.5 proposes coherent intermediate lemmas and drives Lean 4 through each step. Time saved versus writing everything by hand is real: we estimate two to three times faster for an engineer already fluent in Lean.
Second case: property audits on critical Rust code (a concurrent data structure, memory allocation, pool invariants). The model flagged several proofs that were too weak and proposed strengthened versions. Across the 57 repositories Mistral references, the lab claims 47 violated properties and 11 genuine bugs surfaced. Our own experience lines up: the true-detection rate is solid.
Third case: reformalizing in Lean 4 a written O(log n) proof for an AVL tree implementation. It is one of Mistral’s showcase examples. We were able to replicate the proof with minimal human intervention. The autoformalization piece (turning a human-language proof into Lean code that verifies) works better than expected.
Where the model tires: as soon as the statement drifts from classic patterns (exotic topology theorems, proofs that need picking one trick out of five thousand possible), Leanstral 1.5 suggests paths that go nowhere. The 587 of 672 PutnamBench score is excellent, but the 85 misses are exactly the problems demanding real creative reasoning.
Against an open model like MiniMax M3, pushed on general reasoning, Leanstral stands out precisely by its verticality. M3 is far better at discussing a math problem in natural language. Leanstral is far better at producing valid Lean.
Also on Horizon:
- Zuckerberg AI Agents Slower Than Expected at Meta
- Sakana AI: We Tested the Japanese AI Everyone Talks About
- SpaceX AI Phone Prototype Leaks, Musk Denies It All
Our Verdict: Plug It In or Skip It, Depending on Your Job
We recommend Leanstral 1.5 in three precise setups. First is any academic or industrial lab already working in Lean 4. The model saves substantial time and avoids dead-end reasoning. Local install on a serious GPU cluster fits inside a day.
Second setup: a critical security or blockchain vendor auditing code where errors are expensive. Formal verification on key components, baked into CI: Leanstral 1.5 changes the iteration speed. It is not a replacement for human audits, it is a pre-filter that raises the real red flags.
Third setup: a math teacher or trainer preparing formalized exercises for students. Autoformalizing a full course in Lean 4 becomes doable without spending a semester on it. The return on investment is immediate for any university program.
For everyone else, Leanstral 1.5 is overkill. A marketer, a product manager, a transformation consultant will get nothing out of it. Formal verification is not a need of their job, and a general chat like Sonnet 5 or GPT 5.6 will be a thousand times more useful.
The read is simple. Leanstral 1.5 is a precise tool for a precise job. It is not trying to replace a general chat, it lays a brick on a market few players address. At Mistral, the bet is consistent with the vertical open-weight strategy other labs like DeepSeek are also exploring on the inference side.
The real question for Mistral is turning technical success into enterprise adoption. A formal proof certified by Leanstral and verified by Lean 4 could become a standard deliverable at certain critical software vendors within twelve months. That is where the commercial value of the model will actually be decided.
Follow the story on Horizon.
