Mistral released Leanstral, an open-weight model built for theorem proving and code verification, according to a report posted to the company’s GitHub repository on July 4. The model runs at roughly 119 billion parameters and is built on Mistral’s general-purpose coding framework, extended to reason about formal proofs rather than just generate code. A version of the release has been called Leanstral 1.5, described by Mistral as “an open model for proof engineering.”
The distinction between writing code and proving code correct is the entire premise of formal verification, a discipline that predates the current generative-AI wave by decades and has mostly stayed inside compiler research and safety-critical engineering. Large language models are fluent at producing code that looks right. They are far less reliable at establishing that code is right, which is a different task: constructing a machine-checkable proof that a program satisfies a specification, or that a mathematical claim holds under a formal system such as Lean. Leanstral targets that second task directly.
Mistral’s GitHub repository for the project includes a fork of SafeVerify, which the company says it uses for code agent training and as part of its own evaluation stack. That detail matters more than it first appears: it means Mistral is not positioning Leanstral purely as a research artifact, but as infrastructure it already runs internally to check the code its other models and agents produce. An open release of that infrastructure gives outside teams the same checking layer Mistral uses on itself.
Releasing a nearly 120-billion-parameter model as open weight is consistent with Mistral’s broader strategy of competing on openness rather than matching the compute budgets of OpenAI, Google, or Anthropic. A theorem-proving model is a lower-volume, higher-trust use case than a general chatbot: the buyers are formal-methods teams, compiler and hardware verification groups, and engineering organizations running code-generation agents in domains where a wrong answer has real cost. For those teams, an open weight release means the model can run inside an air-gapped or audited environment rather than through an API, which is often a hard requirement for verification tooling.
Teams building or evaluating AI coding agents should treat Leanstral as a candidate verification layer to sit downstream of code generation, not as a replacement for it. Anyone doing formal methods or safety-critical code review should pull LeanstralReport.pdf from mistralai/LeanstralSafeVerify on GitHub this week and test it against an existing proof or verification benchmark before deciding whether it belongs in a production pipeline.
Per Mistral’s LeanstralReport.pdf, posted to the mistralai/LeanstralSafeVerify repository on GitHub on July 4, 2026.
