Mistral AI Releases Leanstral 1.5 119B Parameter Model Open-Sourced Under Apache 2.0

Mistral AI launched the Leanstral 1.5 model around July 2026, with 119 billion total parameters and 6 billion activated parameters, open-sourced under Apache 2.0, along with a free API endpoint leanstral-1-5. The model is optimized for the Lean 4 formal proof language, enabling automatic generation and verification of mathematical proof code.

Fact Restoration

Leanstral 1.5 belongs to the Mistral Small 4 series, adopting a mixture of experts architecture with 128 experts, activating 4 per token. It supports a 256k token context, accepts text and image inputs, and outputs text. The model is updated from the Leanstral-2603 version, with training divided into three stages: intermediate training, supervised fine-tuning, and reinforcement learning using CISPO.

Mechanism Breakdown

The training process constructs two reinforcement learning environments. In the multi-turn environment, the model receives a theorem statement, submits a proof, reads feedback from the Lean compiler, and iterates until success or budget exhaustion. In the code agent environment, the model directly operates the file system, executes bash commands, and calls the Lean language server to obtain real-time information on goals, errors, and types, thereby completing partial proofs, constructing auxiliary lemmas, and handling context compression. Correctness is verified through Mistral's SafeVerify branch.

Leanstral 1.5 improved pass@1 from 21.9 to 28.9 and pass@8 from 31.9 to 43.2 on the FLTEval benchmark. This benchmark is derived from real pull requests from the Fermat's Last Theorem repository.

Industry Impact

For developers, the Apache 2.0 open-source weights and free API lower the barrier to formal verification. Developers can directly download and deploy the weights, or call the leanstral-1-5 API to run proof tasks in local repositories.

For enterprise users, the model provides verifiable code generation capabilities in safety-critical fields such as aviation and healthcare. It has already proven the time complexity of AVL tree insertion and deletion operations and located five previously unreported hidden defects in automated testing, demonstrating its practicality in real engineering environments.

For the competitive landscape, Leanstral 1.5 achieved 100% on both the miniF2F validation and test sets, solved 587 out of 672 challenging problems on PutnamBench, and achieved new highs of 87% on FATE-H and 34% on FATE-X. It surpasses some larger open-source models with a single pass while costing less than one-seventh of Claude Opus 4.6.

Strategic Assessment

Based on existing benchmarks and real repository performance, Leanstral 1.5 is most likely to be rapidly integrated in high-reliability code generation scenarios. Developers can first test its pass rate on specific Lean repositories via the free API before deciding whether to deploy the open-source weights. When selecting models, enterprises should focus on verifying the actual pass rate and iteration cost under their own formal specifications, rather than relying solely on public benchmark scores.

This assessment is based on publicly available FLTEval, PutnamBench, and actual defect discovery data, and is an analysis of future application paths rather than established facts.