Leanstral 1.5: Proof Abundance for All

Mistral AI released Leanstral 1.5, a free Apache-2.0 licensed model with 119B total parameters and 6B active parameters, the Leanstral team said. The release is positioned as a performance upgrade for proof engineering in Lean 4.

On benchmarks, Mistral reports Leanstral 1.5 saturates miniF2F at 100% on both validation and test sets, solves 587 of 672 PutnamBench problems, and reaches state-of-the-art results of 87% on FATE-H and 34% on FATE-X. On PutnamBench, Mistral says it beats Seed-Prover 1.5 at its high setting by 7 problems at about $4 per problem, versus an estimated $300 or more for Seed-Prover, whose high setting uses a budget of 10 H20-days per problem. Mistral notes that higher-ranked provers operate under different conditions, including natural-language proof guidance or much higher cost, such as Aleph Prover at $54-68 per problem.

The model was trained in three stages: mid-training, supervised fine-tuning, and reinforcement learning with CISPO. It uses two RL environments: a multiturn environment where the model submits Lean proofs and refines them using compiler feedback, and a code agent environment where it edits files, runs bash commands, and uses the Lean language server, verified by a fork of SafeVerify. Mistral reports test-time scaling on PutnamBench Pass@8 rising from 44 problems solved at a 50k token budget to 244 at 200k, 493 at 1M, and 587 at 4M; an AVL-tree proof ran over 2.7 million tokens across 22 compactions.

Mistral also open-sourced FLTEval, where it reports Leanstral 1.5 lifts pass@1 from 21.9 to 28.9 and pass@8 from 31.9 to 43.2, surpassing Opus 4.6's 39.6 at one-seventh the cost. In code verification case studies, a pipeline pairing Aeneas (Rust to Lean translation) with Leanstral flagged 47 violated properties across 57 tested repositories, with 11 pointing to genuine bugs and 5 previously unreported on GitHub. One was in the sign function for zigzag decoding in the datrs/varinteger library, where input Std.U64.MAX caused (value + 1) to overflow. Leanstral 1.5 weights are on Hugging Face and it is available as a free API endpoint named leanstral-1-5.