Test Leanstral 1.5 before proof bugs reach production code

Confirmed: Mistral has released Leanstral 1.5, a Lean 4 specialist model for automated theorem proving, autoformalization, and proof-engineering work. It is open sourced under Apache 2.0, available through Hugging Face, and exposed as a free API endpoint for practical testing. The useful question is not whether it is a broad coding model, but whether it can make verified-code workflows cheaper to test.

What changed

Mistral published Leanstral 1.5 on July 2, 2026 as an update to its Lean 4 proof-engineering line. The release page describes a 119B total-parameter sparse model with about 6B active parameters, trained through mid-training, supervised fine-tuning, and reinforcement learning with CISPO.

The model targets Lean 4 tasks where output can be checked by a compiler or verifier. Mistral says it saturates miniF2F, solves 587 of 672 PutnamBench problems, reaches 87% on FATE-H and 34% on FATE-X, and found 5 previously unknown bugs while testing 57 open-source repositories through a code-verification pipeline.

Why this is early

This is early in the practical sense: the announcement is live and official, but teams still need their own repository-level evaluations before relying on it. AI Weekly independently picked up the Labs availability and model-card details, while Mistral's March Leanstral post gives the baseline for the original release.

The signal is stronger than a rumor because the primary source is Mistral and the availability claim is corroborated by independent coverage. It is still not a production guarantee. Formal methods are domain-specific, and benchmark wins do not prove that your proof harness, library versions, or repository style will work cleanly.

Key takeaways

• Leanstral 1.5 is built for Lean 4 proof work, not general chat.
• Mistral says it is Apache-2.0 licensed, open sourced, available on Hugging Face, and reachable through a free API endpoint.
• The strongest workflow fit is theorem proving, autoformalization, and verified-code checks where Lean can validate the output.
• The release claims real repository testing, including previously unknown bugs found through a Rust-to-Lean verification pipeline.
• Teams should run internal evals before replacing existing proof, fuzzing, or review workflows.

Availability and access

Mistral says Leanstral 1.5 is available now through Hugging Face and as the leanstral-1-5 API endpoint. The company recommends using it in Mistral Vibe, with Lean LSP MCP configured for proof work.

Pricing needs a careful read. Mistral calls the endpoint free in the announcement, and AI Weekly reports the Labs listing at $0. Treat that as launch availability, not a permanent pricing contract. Check the current Mistral console before planning a long-running proof pipeline around free usage.

Practical LinkLoot angle

Leanstral 1.5 is useful when a team wants AI help on code that must be proven, not merely reviewed. A practical test loop is simple: pick a small Lean repository, choose failing or incomplete proofs, run Leanstral with the same compiler version your team uses, then score compile success, patch quality, runtime cost, and reviewer time saved.

Do not frame it as a replacement for fuzzing, tests, or human review. Use it as another verification lane. For teams building coding-agent workflows, this belongs next to LinkLoot's AI workflow automation guide: the model is strongest when the workflow has a verifier, a clear task budget, and a repeatable acceptance test.

What to verify before you act

• Confirm the current Mistral API endpoint name, rate limits, and pricing in your account.
• Check the Hugging Face model page for files, license text, and serving notes before self-hosting.
• Run the model against your Lean version, dependencies, and repository structure.
• Compare against your current proof assistant workflow using compile success, reviewer time, and false-positive bug reports.
• Review data handling before sending proprietary code or specs through hosted APIs.

Source check

Confirmed by: Mistral's Leanstral 1.5 announcement says the model is released, open sourced, available via Hugging Face, and exposed as a free API endpoint. It also gives the main benchmark and repository-testing claims.

Early signal / context: AI Weekly independently reports the Labs availability, parameter profile, and $0 listing. Mistral's original Leanstral announcement provides context for the March baseline and the model family's formal-verification focus.

LinkLoot will treat updated pricing, changed API availability, fresh model-card files, or independent benchmark results as update triggers rather than rewriting launch claims as production proof.