Leanstral 1.5 Solves 587 PutnamBench Problems, Finds 5 New Bugs

Q: Why is a breakthrough AI model being retired just three months after its high-profile launch?

It is being retired because Mistral has already replaced the original Leanstral with a stronger successor, Leanstral 1.5, not because the project failed. The original Leanstral launched in March 2026 and was discontinued when Leanstral 1.5 was released on June 30, 2026. The newer model delivers a major jump in formal-verification performance, including perfect results on miniF2F, 587 solved problems on PutnamBench, and record scores on FATE-H and FATE-X. It also showed practical value by finding previously unknown bugs in open-source code and handling long, complex proof workflows in Lean 4. The searched material also shows that Leanstral 1.5 itself has a short published lifecycle: Mistral says it will be retired on September 30, 2026. No official reason is given for that date. The most likely explanation is product-cycle management: Mistral appears to be iterating quickly, replacing models with newer versions rather than maintaining many overlapping endpoints. There is no evidence in the provided material that the retirement was driven by a scandal, export restriction, or a collapse in technical quality. If anything, the evidence points the other way: Leanstral 1.5 is high-performing, open-source, and inexpensive to run. The only notable technical caveat in outside evaluation was weaker generalization on the miniCTX benchmark, but nothing suggests that this is why it is being retired.

3 articles · Updated · Mistral AI · Jul 2

A free Apache-2.0 Lean 4 model with 6B active parameters posted a major formal-verification jump, saturating miniF2F at 100% and setting new highs of 87% on FATE-H and 34% on FATE-X.
587 of 672 PutnamBench problems were solved after training through mid-training, supervised fine-tuning and CISPO-based reinforcement learning, with the model showing steady gains as token budgets rose to 4 million.
In practical proof engineering, Leanstral 1.5 lifted FLTEval pass@1 to 28.9 and pass@8 to 43.2, beating Opus 4.6's 39.6 at about one-seventh the cost.
Across 57 open-source repositories, an automated Rust-to-Lean verification pipeline flagged 47 violated properties and confirmed 11 real bugs, including 5 previously unreported issues.
Hugging Face weights, a free API and open-sourced FLTEval make the release broadly accessible, underscoring a push to move formal verification from benchmarks into real-world Lean 4 development.

Sources

Mistral AI2d ago

Leanstral 1.5 Released: State-of-the-Art Formal Verification for Lean 4

MarkTechPost2d ago

Mistral AI Releases Leanstral 1.5: An Apache-2.0 Lean 4 Code Agent Model Solving 587 of 672 PutnamBench Problems - MarkTechPost

7minai.com7h ago

How to Run Leanstral 1.5 for Lean 4 Proofs (Local vLLM and the Free API Shortcut)

If an AI aces benchmarks but fails at generalization, can we trust it with our critical code?

Why is a breakthrough AI model being retired just three months after its high-profile launch?

Do free AI provers signal innovation for all or the bursting of a big tech 'inference bubble'?

Leanstral 1.5 Launches with 119B Parameters and Free Labs Access, Advancing Formal Verification for All

Overview

Leanstral 1.5, released on July 1, 2026, marks a significant update in AI-assisted formal verification, but immediate confirmation of its performance improvements and open-source details remains limited. The model card does not yet provide new benchmarks since the March version, and publications like AI Weekly have not confirmed any new comparison results or policies about releasing the model’s weights. As a result, comprehensive data on Leanstral 1.5’s advancements and its full accessibility are still awaiting official disclosure, highlighting the need for further transparency before its impact can be fully assessed.

...