Mistral releases Leanstral 1.5, an open-weights Lean 4 proof agent that saturates miniF2F and solves 587/672 PutnamBench problems
Confirmed
The takeaway
On July 2, 2026 Mistral released Leanstral 1.5 (“Proof abundance for all”), an Apache-2.0, 119B-parameter (6B active) MoE model specialised for Lean 4 theorem proving and code verification.
Status
- Claim
Confirmed
- Our reporting
- High confidence
- Importance
- 3 of 5
- Last verified
- 1 October 2026
Your AI and this story
- GPT-6.2 Astra32 days after its cutoff
- Claude Opus 5.52 days after its cutoff
- Gemini 4 Hydrogen93 days after its cutoff
- Grok 4.732 days after its cutoff
None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 2 days before it.
Key facts
- Released July 2, 2026; 119B total / ~6B active parameters (MoE, 128 experts, 4 active per token); Apache 2.0; 256K context (≤200K recommended)
- Benchmarks (Mistral): miniF2F 100% (validation and test); PutnamBench 587/672; FATE-H 87%; FATE-X 34%; FLTEval pass@1 28.9%, pass@8 43.2%
- Test-time scaling on PutnamBench: 44 problems solved at a 50k-token budget, rising monotonically to 587 at 4M tokens
- Training: mid-training, SFT, then RL with CISPO in two environments (multi-turn theorem proving and code-agent workflows)
- Code verification: found 5 previously unreported bugs across 57 open-source repositories, including an integer overflow in datrs/varinteger’s zigzag decoding
- Predecessor: Leanstral (Leanstral-2603), released March 16, 2026 as ‘the first open-source code agent for Lean 4’ (Mistral)
- Access: Hugging Face mistralai/Leanstral-1.5-119B-A6B; free Mistral API endpoint, model id leanstral-1-5
What happened
Mistral published Leanstral 1.5 on July 2, 2026. It is a sparse MoE model trained to work inside real Lean 4 repositories, writing proofs as well as formal specifications and proofs about software. The release is an update of Leanstral (March 2026), which Mistral had billed as the first open-source code agent for Lean 4. The headline numbers are a saturated miniF2F (100%) and 587/672 on PutnamBench, with performance that keeps rising as the token budget grows to 4M tokens. Mistral also ran it as a code-verification agent over 57 open-source repositories and reports 5 new bugs.
Why it matters
Formal proof was a frontier-lab capability (AlphaProof, Aristotle, Seed-Prover) through 2025. By mid-2026 an Apache-licensed model with only ~6B active parameters reports near-saturation of the standard competition benchmarks, and it is free to call. That makes Lean verification cheap enough to bolt onto the summer-2026 wave of AI-generated mathematics and code. Benchmark numbers are Mistral’s own and were not independently checked here.
Sources
4 sources from 4 sites. Numbers match the chips in the text.
4 sources: 3 primary, 1 press
Primary
- Mistral AI: Leanstral 1.5, proof abundance for allmistral.ai, official
- Hugging Face: mistralai/Leanstral-1.5-119B-A6Bhuggingface.co, code
- Mistral AI for Developers on X: Leanstral release (March 16, 2026)x.com, official
Press
- NYU Shanghai RITS: Leanstral, Mistral’s open-source proof agent for Lean 4rits.shanghai.nyu.edu, press
Changes
- Filed