@mertunsal2020: Today, we are releasing Le Chaton L∃∀N, aka Leanstral 1.5. It achieves SOTA performance on graduate algebra benchmarks …
Summary
Today, we are releasing Le Chaton L∃∀N, aka Leanstral 1.5, which achieves SOTA performance on graduate algebra benchmarks FATE-H and FATE-X and improves the Pareto Frontier on PutnamBench, solving 587/672 problems with a x10 cheaper budget.
View Cached Full Text
Cached at: 07/04/26, 08:55 PM
Today, we are releasing Le Chaton L∃∀N, aka Leanstral 1.5.
It achieves SOTA performance on graduate algebra benchmarks FATE-H and FATE-X and improves Pareto Frontier on PutnamBench, solving 587/672 problems with a x10 cheaper budget.
Leanstral 1.5 shows the strongest test-time scaling we have seen from a formal-reasoning model. The figure below tracks Pass@8 on PutnamBench as we raise the token budget per attempt from 25k to 4M: performance climbs smoothly the whole way.
Despite being primarily trained on math, Leanstral demonstrates impressive code verification capabilities, discovering previously unknown bugs in open-source repositories.
We built an automated pipeline where Aeneas translates Rust code to Lean and Leanstral infers the user intent and generates correctness properties from the code. Across 57 tested repositories, this process flagged 47 violated properties, with 11 pointing to genuine bugs—5 of them previously unreported on GitHub.
We believe every mathematical breakthrough should be automatically verified in Lean.
That means formalizing all of mathematics—and this will require us to run agents at an extremely large scale, perhaps burning through billions of tokens. We hope that Leanstral can be the foundation to get there.
Paper: https://github.com/mistralai/LeanstralSafeVerify/blob/main/LeanstralReport.pdf… Blog Post: https://mistral.ai/news/leanstral-1-5/… Model: https://huggingface.co/mistralai/Leanstral-1.5-119B-A6B…
it’s by compaction, see tech report!
TY! Not hating on Mistral for once
ty Alex - I am in PA we should meet
It is benchmaxxed indeed, but in a good way for formal! that’s the point of a specialized model
Similar Articles
@sophiamyang: Introducing Leanstral 1.5 A 119B (6B active) open model for formal proof engineering in Lean 4: 100% on miniF2F 587/672…
Introducing Leanstral 1.5, a 119B parameter (6B active) open model for formal proof engineering in Lean 4, achieving 100% on miniF2F, state-of-the-art scores on PutnamBench and FATE benchmarks, and discovering previously unknown bugs in open-source repositories.
Leanstral 1.5: Proof Abundance for All
Mistral AI releases Leanstral 1.5, a 6B active parameter model for Lean 4 proof engineering, achieving state-of-the-art results on multiple formal verification benchmarks and uncovering real-world bugs, fully open-sourced under Apache-2.0.
Leanstral 1.5
Mistral AI releases Leanstral 1.5, an updated Lean 4 formal proof engineering model optimized for automated theorem proving and autoformalization, with 119B total parameters and 6.5B active parameters.
@rohanpaul_ai: atomic[.]chat, a desktop app that runs LLMs locally, ran a very revealing comparison for Claude Sonnet 5, Claude Opus 4…
atomic.chat ran a comparison showing Claude Sonnet 5 matches GPT 5.5 on three physics coding demos at 6x lower cost, using fewer tokens than other models.
@leloykun: [WIP] Blog post on Lean4-to-TileLang Tensor Program Superoptimizer here:
A technical blog post introduces a Lean4-to-TileLang tensor program superoptimizer that automatically generates optimized GPU/TPU kernels and hyperparameter scaling laws, demonstrating performance gains over torch.compile.