Leanstral 1.5

// Overview

Leanstral 1.5 is an open-weight code agent model from Mistral AI, released on 2 July 2026 and specialized for Lean 4 formal proof engineering. It succeeds the original Leanstral (Leanstral-2603) with substantial gains on automated theorem proving, autoformalization, and code verification, and is distributed under the permissive Apache 2.0 license on Hugging Face (mistralai/Leanstral-1.5-119B-A6B).

The model is a mixture-of-experts with 119B total parameters and roughly 6.5B active per token (128 experts, 4 active), built on the Mistral Small 4 family. It accepts text and image inputs and supports a 256K-token context window (≤200K recommended for deployment). Training combines mid-training on Lean-focused data, supervised fine-tuning, and reinforcement learning with the CISPO objective, together with a test-time scaling recipe that lets the model spend more compute on hard proofs.

On the standard formal-mathematics benchmarks, Leanstral 1.5 saturates miniF2F (100%), solves 587 of 672 PutnamBench problems, and reports 87% on FATE-H and 34% on FATE-X. On FLTEval (formal Lean tactics) it reaches 28.9 pass@1 and 43.2 pass@8. Beyond theorem proving, Mistral reports the model uncovered five previously-unknown bugs in a survey of 57 open-source repositories, demonstrating a working agentic pipeline for real-world code verification. It is available as the free `leanstral-1-5` endpoint in Mistral Labs and integrated with the Vibe coding surface.

// Benchmarks

1. miniF2F100%
2. PutnamBench (problems solved)587 / 672
3. FATE-H87%
4. FATE-X34%
5. FLTEval pass@128.9%
6. FLTEval pass@843.2%

Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.

// Strengths

• Saturates miniF2F (100%) and reports state-of-the-art 587/672 on PutnamBench for formal math
• Apache 2.0 open weights on Hugging Face, free to self-host and fine-tune commercially
• MoE efficiency: 119B total but only ~6.5B parameters active per token
• 256K-token context window with a test-time scaling recipe for harder proofs
• Practical Lean 4 code-verification agent — surfaced five previously-unknown bugs across 57 open repos

// Best for

• Reach for it when you need an open model for Lean 4 theorem proving or autoformalization work.
• Reach for it to run local or self-hosted formal-verification agents over Lean 4 codebases.
• Reach for it as a starting point for research on RL for formal math (CISPO training recipe).
• Reach for it in the free Mistral Labs endpoint to prototype proof-engineering agents at zero cost.

// How to access