Breaking

Mistral Releases Open Model Leanstral 1.5 for Lean 4 Proof Work

July 4, 2026 at 08:57 EDT

Foundation Models
Open Source
Software Dev & Coding

On June 30, 2026, Mistral AI released Leanstral 1.5 (formally Leanstral-1.5-119B-A6B), an open-weight model specialized in formal verification for Lean 4, the proof assistant used to mechanically check mathematical proofs. The weights are hosted on Hugging Face under an Apache-2.0 license and it is also available via a free API.

July 2026 · Mistral AI · Apache-2.0

Leanstral 1.5: Proof Abundance for All

An open-weight Lean 4 code agent that saturates miniF2F, solves 587 of 672 PutnamBench problems, and turns raw token budget directly into verified mathematical proofs — free on Hugging Face with a free API.

100%

miniF2F saturated (valid & test)

587/672

PutnamBench solved (Pass@8)

119B/6B

Total / active params (MoE), 256k context

~$4

Cost per PutnamBench problem

Benchmark scores

Percent solved — taller column = higher score

100%

87%

43%

34%

miniF2F

Putnam
(587/672)

FATE-H
SOTA

FLTEval
pass@8

FATE-X
SOTA

The verified proof loop

Generate
Lean 4 proof

→

Compiler judges
as verifier

→

Spend more
token budget

↺

Performance rises monotonically with token budget — one AVL-tree complexity proof consumed ~2.7M tokens and 22 compactions. In a Rust→Lean pipeline it found 5 previously unreported bugs across 57 repositories.

Strengths

Open weights, Apache-2.0, free API & Vibe integration
Surpasses Seed-Prover 1.5 high on Putnam without NL guidance
Token budget converts directly into verified results

Limits

Specialized for math/proofs — poor at general coding or chat
Local inference needs quantization; ~9 tokens/sec reported
Hard proofs demand long time & heavy token consumption

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.