Mistral AI releases Leanstral 1.5, a 119B MoE model for Lean 4 formal proof engineering · Digg

/Tech14h ago

Mistral AI releases Leanstral 1.5, a 119B MoE model for Lean 4 formal proof engineering

Story Overview

Mistral just dropped an open 119B-parameter MoE model tuned specifically for Lean 4 formal proof work, where only about 6B parameters stay active at once. It saturates the miniF2F benchmark at 100 percent while posting 587 solved problems on PutnamBench and fresh highs on the FATE-H and FATE-X sets.

1712.1K148640303.3K

Original post

kache@yacineMTB#403inTech

Awesome stuff The ability to produce narrow models is the product

Albert Jiang@AlbertQJiang

https://mistral.ai/news/leanstral-1-5/ Leanstral 1.5 is here. SoTA on FATE-H/X, 587 on PutnamBench, saturating miniF2F, all with an Apache-2 6B active params model. We are having fun verifying code properties and catching bugs in Rust repos!

Tech report covering training environment and evaluations: https://github.com/mistralai/LeanstralSafeVerify/blob/main/LeanstralReport.pdf We also open-source LeanstralSafeVerify and FLTEval.

11:41 AM · Jul 3, 2026 · 9.3K Views

Developer Impact

It ships ready for agentic proof loops

The release includes support for multi-turn workflows, filesystem navigation, and Lean language server hooks, so developers can drop it straight into code-verification agents without extra scaffolding.

Access Now

Weights and a free API endpoint are live today

Everything is Apache-2.0 on Hugging Face plus a labs-tier endpoint, though longer-term pricing or enterprise hosting details remain unspecified in the announcement.

Sentiment

Positive users praise Mistral's Leanstral 1.5 for its SOTA PutnamBench results and open 119B Lean 4 model, while negative users call the benchmarks saturated or benchmaxxed and dismiss the release as VC hype.

Pos

84.1%

Neg

15.9%

58 comments with sentiment.

Cluster Engagement

Digg Deeper

No Digg Deeper questions have been answered for this story yet.

Related links

LeanstralSafeVerify/LeanstralReport.pdf at main · mistralai/LeanstralSafeVerify

GITHUBVia

Leanstral 1.5: Proof Abundance for All

MISTRAL AI

Posts from X

Most Activity

VIEWS13.1KBOOKMARKS79LIKES137REPLIES5

elie@eliebakouch

very nice release by mistral (with a tech report!!), they replace the complicated scaffold of previous lean models by a simple coding agent and just scale test time compute budget with compaction every 200k

it's also quite interesting that you don't really see the impact of compaction in this curve (or maybe this is why they see diminishing returns?)

~6B lean mid training token, SFT with 50% lean traces and then RL with CISPO

Mert Ünsal@mertunsal2020

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.

🧵

9h13.1K13779

RETWEETS63

Mert Ünsal@mertunsal2020

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.

🧵

19h166.3K1.1K293

Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞)@teortaxesTex

I'd say ≈5 months behind Seed but this is good

Albert Jiang@AlbertQJiang

https://mistral.ai/news/leanstral-1-5/ Leanstral 1.5 is here. SoTA on FATE-H/X, 587 on PutnamBench, saturating miniF2F, all with an Apache-2 6B active params model. We are having fun verifying code properties and catching bugs in Rust repos!

Tech report covering training environment and evaluations: https://github.com/mistralai/LeanstralSafeVerify/blob/main/LeanstralReport.pdf We also open-source LeanstralSafeVerify and FLTEval.

11h5K5114

Beff (e/acc)@beffjezos

>Le Chaton Lean release implies the existence of Le Chaton Fat 🧐

Mert Ünsal@mertunsal2020

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.

🧵

7h8.8K745

Mert Ünsal@mertunsal2020

🗞️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

19h4.6K579

Mert Ünsal@mertunsal2020

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.

19h5.4K551

Mert Ünsal@mertunsal2020

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.

19h5K431

Mert Ünsal@mertunsal2020

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.

19h5.1K45

Levy@_levyathan_

@mertunsal2020 I know you guys are hiding Le Chaton Fat

15h1.1K18

Kfir Gollan@kfirgollan

@mertunsal2020 @grok why Fable isnt on these benchmarks? How it compares on these math problems?

15h3.4K9

jkazi@JawadKazi13

@mertunsal2020 How did you manage context at 4m? Could you have gone to 10m possibly getting a score >600 👀

17h3K12

elie@eliebakouch

Albert Jiang@AlbertQJiang

https://mistral.ai/news/leanstral-1-5/ Leanstral 1.5 is here. SoTA on FATE-H/X, 587 on PutnamBench, saturating miniF2F, all with an Apache-2 6B active params model. We are having fun verifying code properties and catching bugs in Rust repos!

Tech report covering training environment and evaluations: https://github.com/mistralai/LeanstralSafeVerify/blob/main/LeanstralReport.pdf We also open-source LeanstralSafeVerify and FLTEval.

9h85731

Amïn@Shoier__

@mertunsal2020 Either "fat" or nothing.

16h1.2K10

Dominic@dommcdev

@mertunsal2020 3 benchmarks I've never heard of featuring a small lead above models I've never heard of...

13h50510

Arjun Rawat@ArjunRawatCell

@mertunsal2020 nobody cares give us le chaton fat (jokes)

16h1.5K9

Erik@LilDombi

@kfirgollan @mertunsal2020 @grok Huge expensive generalized model vs tiny cheap specialized model.

15h6538

elie@eliebakouch

cc @mertunsal2020 @AlbertQJiang did you see some limitation with compaction at 200k when scaling to 4M actually?

elie@eliebakouch

very nice release by mistral (with a tech report!!), they replace the complicated scaffold of previous lean models by a simple coding agent and just scale test time compute budget with compaction every 200k

it's also quite interesting that you don't really see the impact of compaction in this curve (or maybe this is why they see diminishing returns?)

~6B lean mid training token, SFT with 50% lean traces and then RL with CISPO

7h58440

Rota 🚪🧎‍♂️@pli_cachete

@mertunsal2020 @AlbertQJiang Fantastic

18h1.5K7

BenIt Pro@BennettBuhner

@mertunsal2020 Oh that's cool! But wen SOTA in SWE? 👀

15h1.1K7

Albert Jiang@AlbertQJiang

@noctus91 Oh yes

16h3263