Mistral AI Releases Leanstral 1.5: Free AI Model Sets Formal Verification Records

Mistral AI on July 2, 2026, presented Leanstral 1.5 — an updated model for formal verification with 6 billion active parameters, distributed freely under the Apache 2.0 license. The model completely saturated the miniF2F benchmark, solved 587 out of 672 PutnamBench tasks, 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 language and code verification in real repositories. Key results of the release:

• 100% on miniF2F — complete saturation on both validation and test sets
• 587 out of 672 PutnamBench tasks — 7 more than competitor Seed-Prover 1.5
• 87% on FATE-H and 34% on FATE-X — new records on graduate and doctoral level tasks
• Around $4 per task compared to $300 for Seed-Prover 1.5 in high mode
• 5 previously unknown bugs found in 57 open repositories

The model was separately tested on the FLTEval benchmark, based on real pull requests from the Great Fermat's Last Theorem repository — confirming applicability to industrial-scale tasks.

How Leanstral 1.5 was trained

Leanstral 1.5 went through three preparation stages: pre-training, supervised learning, and reinforcement learning using 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 success or computational budget exhaustion.

In the agent environment, the model works like a developer in a real file system: edits files, executes bash commands, and accesses the Lean language server in real time to inspect goals, errors, and type information. This mode enables solving long tasks: completing partial proofs in repositories, building auxiliary lemmas, and preserving progress across multiple rounds of context compression.

Why this matters

Formal verification is one of the most demanding tasks for language models: success requires mathematical perfection and long chains of reasoning. Powerful solutions in this area have previously been either costly or closed.

Leanstral 1.5 changes this equation. A Mixture of Experts architecture with 119 billion total and only 6 billion active parameters makes inference economical. The Apache 2.0 license permits commercial use without restrictions.

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

Leanstral team at Mistral AI states.

What it means

Mistral AI is opening access to a formal verification tool that outperforms closed competitors on price-to-quality ratio. Publication on Hugging Face and a free API lower the entry barrier for researchers and developers and can accelerate the penetration of formal methods into industrial software development.

Frequently asked questions

Where can I get access to Leanstral 1.5?

The model is published on Hugging Face under the Apache 2.0 license and available through Mistral AI's free API. Commercial use is permitted without additional conditions.

How much does working with Leanstral 1.5 cost?

About $4 per proof of one task. Seed-Prover 1.5 in high mode consumes 10 GPU-days on H20 processors and costs approximately $300 per task — a 75-fold difference.