AI/TLDR

Mistral AI · 2026-07-02 · major

Leanstral 1.5 — Mistral's updated Lean 4 formal-proof model

Leanstral 1.5 saturates miniF2F at 100%, solves 587/672 PutnamBench problems, and finds 5 previously unreported bugs across 57 open-source repos. Apache-2.0 weights on Hugging Face and a free labs-leanstral-1-5 API.

Mistral AI announcement banner for Leanstral 1.5 — Proof Abundance for All

Mistral's Lean 4 theorem-prover saturates miniF2F, tops PutnamBench, and ships as an Apache-2.0 119B/6.5B MoE.

Key specs

Parameters119B
Active params6.5B
Mini f2 f100%
Putnam bench587/672

Quick facts

MakerMistral AI
Version1.5 (Jul 2, 2026)
Params119B total / 6.5B active (MoE)
LicenseApache-2.0
AccessOpen weights on Hugging Face + free API endpoint labs-leanstral-1-5
Context window256K tokens
FocusLean 4 theorem proving and autoformalization

Benchmarks

miniF2F (validation + test)
Leanstral 1.5100%
source ↗
PutnamBench (problems solved / 672)
Leanstral 1.5587
Seed-Prover 1.5580
source ↗
FATE-H
Leanstral 1.587%
source ↗
FLTEval Pass@1
Leanstral 1.528.9%
Claude Opus 4.639.6%
source ↗

What is it?

Leanstral 1.5 is Mistral's mixture-of-experts model tuned for Lean 4 formal proofs and autoformalization — writing machine-checkable proofs and turning informal math into Lean statements. Released under Apache-2.0 with 119B total and 6.5B active parameters, weights are on Hugging Face and the same model runs for free as labs-leanstral-1-5 on Mistral's API.

How does it work?

A sparse expert routing keeps compute low while giving the model access to a wide bank of specialised weights for tactics, term-mode edits, and autoformalization patterns. Training combined mid-training, supervised fine-tuning, and reinforcement learning with CISPO across two environments: multi-turn Lean proof interaction and a code-agent workflow that reads repos, writes properties, and drives the Lean kernel.

Why does it matter?

Formal proof is one of the few coding tasks with a hard ground-truth checker, so a stronger prover directly reduces verification busywork. Leanstral 1.5 saturates miniF2F, solves 587/672 PutnamBench problems, and in a 57-repo sweep uncovered 5 previously unreported bugs — including an integer-overflow in datrs/varinteger — at roughly $4 per PutnamBench problem versus $300+ elsewhere.

Who is it for?

Formal-methods researchers, Lean 4 users, and teams building verified libraries or protocols.

Frequently asked questions

Is Leanstral 1.5 open source?
Yes. Leanstral 1.5 ships under Apache-2.0 with full weights hosted at huggingface.co/mistralai/Leanstral-1.5-119B-A6B. Anyone can download, fine-tune, and self-host the 119B-parameter MoE, and Mistral also runs a free API endpoint at model id labs-leanstral-1-5 for users who prefer not to run it themselves.
How much does Leanstral 1.5 cost to use?
Leanstral 1.5 is free through Mistral's Labs endpoint (model id labs-leanstral-1-5). Mistral says a full PutnamBench proof search costs about $4 per problem on their infrastructure, versus an estimated $300+ per problem on comparable frontier models — the pricing gap is central to the 'proof abundance' framing.
Which benchmarks does Leanstral 1.5 lead on?
Leanstral 1.5 saturates miniF2F at 100% on both validation and test sets, solves 587 of 672 PutnamBench problems (seven more than Seed-Prover 1.5), and reaches 87% on FATE-H. On FLTEval it lands Pass@1 28.9% and Pass@8 43.2%, competitive with Claude Opus 4.6 at roughly one-seventh the compute cost.
Has Leanstral 1.5 found real bugs in open-source code?
Yes. Applied to 57 Rust and Go repositories, Leanstral 1.5 flagged 47 violated properties, of which 11 were genuine bugs and 5 were previously unreported on GitHub. One concrete find: an integer-overflow crash in the datrs/varinteger zigzag-decoding sign function on input Std.U64.MAX.
How is Leanstral 1.5 different from the June model card release?
Leanstral 1.5 first appeared as a Labs model card on June 30, 2026. The July 2 'Proof Abundance for All' announcement is the same 119B/6.5B MoE with the full benchmark disclosure, real-world bug-hunting results, and an Apache-2.0 open-weights drop on Hugging Face — turning a quiet playground preview into a public open-source release.

Try it

API: model id labs-leanstral-1-5. Weights: https://huggingface.co/mistralai/Leanstral-1.5-119B-A6B.

Sources · 3 outlets

Tags

  • mistral
  • leanstral
  • lean-4
  • theorem-proving
  • autoformalization
  • formal-methods
  • verified-software
  • moe
  • apache-2-0
  • open-weights

← All releases · Learn AI