Mistral's Leanstral 1.5 targets Lean 4 theorem proving with a sparse 119B MoE

Mistral quietly added Leanstral 1.5 to its experimental Labs line on June 30, 2026. Unlike the company’s general-purpose chat models, this one is narrowly specialized for formal mathematics: generating and verifying proofs in Lean 4, and autoformalization — the task of translating informal mathematical statements into machine-checkable form. It’s a tool for researchers in formal verification, not a consumer assistant.

The architecture is a sparse mixture-of-experts: 119 billion total parameters but only 6.5 billion active per token, so inference costs track a small model while the network retains far larger capacity — the likely reason for the ‘Lean’ branding. It ships with a 256k-token context window (useful for holding large proof states and library context), plus function calling, structured outputs, and fill-in-the-middle support. It’s served through Mistral’s API as labs-leanstral-1-5 and is currently free to use.

Domain-specialized theorem provers are an active research frontier, with direct relevance to verified software and automated mathematics, so a dedicated, efficiently-served model at zero cost is notable. That said, the model card is thin: Mistral publishes no benchmark results and no licensing or open-weights details, and the ‘Labs’ designation flags this as experimental — making it hard to judge how Leanstral 1.5 stacks up against competing proof models on capability.