Mistral Releases Open-Source Lean 4 Proving Agent Leanstral 1.5

Mistral AI has released Leanstral 1.5, an open-source code agent model specifically designed for Lean 4 formal proofs and theorem proving. The model features 119B parameters with 6B activated and is released under the Apache-2.0 license, aiming to deeply integrate LLM capabilities with formal mathematical verification to automate mathematical reasoning.

Key Details and Performance

Leanstral 1.5 achieves state-of-the-art (SOTA) results across several authoritative math benchmarks. According to Sophia Yang, the model scored 100% on the miniF2F benchmark. In the PutnamBench test, it successfully solved 587 out of 672 problems (at an average cost of about $4 per problem). Additionally, it achieved 80% accuracy on the FATE-H test and secured SOTA performance on a graduate-level algebra benchmark.

Working Principle and Application Value

In her analysis, Sophia Yang highlighted that Leanstral's core value lies in distinguishing between traditional "testing" and "proving." While testing only verifies specific inputs, proofs written in Lean and checked by a computer can assert that a property holds for all inputs. Acting as a proof-writing system, Leanstral can autonomously edit proof code, read checker error messages, and iteratively retry, significantly enhancing the efficiency of automating formal proofs.

2026-07-03 ~ 2026-07-05 · 5 related posts

• Leanstral 1.5 Released: SOTA on Grad Algebra Benchmark — AlbertQJiang · 2026-07-03
• [source] Leanstral 1.5 Open-Sourced for Lean 4 — sophiamyang · 2026-07-04
• Mistral Releases Leanstral 1.5: Open-Source Lean 4 Agent on Apache — sophiamyang · 2026-07-04
• Mistral Unveils Leanstral: First Open-Source Agent for Lean 4 — mervenoyann · 2026-07-04
• [source] Explaining Mistral's Formal Proof System Leanstral — sophiamyang · 2026-07-05