Mistral AI releases Leanstral 1.5: a free AI model that sets formal verification records

Mistral AI on July 2, 2026 introduced Leanstral 1.5, an updated formal verification model with 6 billion active parameters, distributed for free under the Apache 2.0 license. The model fully saturated the miniF2F benchmark, solved 587 of 672 PutnamBench problems, and set records on FATE-H (87%) and FATE-X (34%).

What Leanstral 1.5 can do

The model specializes in formal theorem proving in Lean 4 and in verifying code in real repositories. Key results of the release:

• 100% on miniF2F, full saturation on both splits (validation and test)
• 587 of 672 PutnamBench problems, 7 more than competitor Seed-Prover 1.5
• 87% on FATE-H and 34% on FATE-X, new records on graduate-level and doctoral-level problems
• About $4 per problem, versus $300 for Seed-Prover 1.5 in high mode
• 5 previously unknown bugs found across 57 open-source repositories

The model was also tested separately on the FLTEval benchmark, which is built from real pull requests in the repository for the proof of Fermat's Last Theorem. This confirms that it applies to industrial-scale tasks.

How the training works

Leanstral 1.5 went through three training stages: continued pretraining, supervised training, and reinforcement learning with the CISPO method. In the final stage, the model trained in two environments.

In the multi-step environment, Leanstral receives a theorem statement, submits a proof attempt, gets feedback from the Lean compiler, and iterates until it succeeds or exhausts its compute budget.

In the agentic environment, the model works like a developer in a real file system: it edits files, runs bash commands, and queries the Lean language server in real time to inspect goals, errors, and type information. This mode lets it handle long tasks: completing partial proofs in repositories, building auxiliary lemmas, and preserving progress across several rounds of context compaction.

Why it matters

Formal verification is one of the most demanding tasks for language models: success requires mathematical flawlessness and long chains of reasoning. Powerful solutions in this area were previously either expensive or closed. Leanstral 1.5 changes that balance.

The Mixture of Experts architecture, with 119 billion total and only 6 billion active parameters, keeps inference economical. The Apache 2.0 license permits unrestricted commercial use.

"Rigorous formal methods can be both effective and practical for real-world use," the

Leanstral team at Mistral AI said.

What this means

Mistral AI is opening access to a formal verification tool that beats closed competitors on price-to-quality ratio. Publication on Hugging Face and a free API lower the barrier to entry for researchers and developers, and may speed the adoption of formal methods in industrial software development.

Frequently asked questions

Where can I get Leanstral 1.5?

The model is published on Hugging Face under the Apache 2.0 license and is available through Mistral AI's free API. Commercial use is allowed with no additional conditions.

How much does it cost to use Leanstral 1.5?

About $4 per proof of a single problem. Seed-Prover 1.5 in high mode uses 10 GPU-days on H20 processors and costs roughly $300 per problem, a 75x difference.