Mistral formal proof model · checked July 30, 2026 Apache 2.0 · Public Preview API

Quick answer

Leanstral 1.5 works inside the proof loop.

It is a specialized Lean 4 model with Apache-2.0 weights and a public-preview Mistral API. Use compiler feedback, inspect every assumption and edit, and reproduce the result from clean state before calling a proof verified.

Verified provider record

Purpose

Lean 4 formal proof engineering, automated theorem proving, autoformalization, and repository-scale proof work.

Architecture

Mistral's release says 119B total and 6B active parameters; the current model card says 119B total and 6.5B active. Preserve that source difference.

License and weights

Apache 2.0, with downloadable weights linked by Mistral from the release post.

Current API record

The model card checked July 30, 2026 lists Public Preview, $0 pricing, 256k context, 128k maximum output, and the alias labs-leanstral-1-5.

Release-page API record

The July 2 release calls the endpoint free and names leanstral-1-5. Verify the live model catalog before coding because it differs from the current model card.

Flowith access

Not established. Mistral weights or API availability does not prove a Flowith model mapping.

Proof engineering

Make verification an observable workflow

Step 1

State the proof target

Pin the repository revision, Lean and mathlib versions, theorem statement, allowed assumptions, target files, and acceptance conditions.

Step 2

Give tool feedback

Use Lean compiler and language-server feedback in the loop. A plausible proof string is not a verified proof.

Step 3

Review the proof boundary

Check for changed statements, added axioms, unsafe declarations, admitted results, dependency drift, and edits outside the approved scope.

Step 4

Reproduce from clean state

Compile in a clean environment, run repository checks, inspect the diff, and retain a human proof-engineering review.

Read benchmarks with their setup

  • Mistral reports 100% on both miniF2F validation and test. This is a provider-published evaluation, not an independent guarantee for a new repository.
  • Mistral reports 587 of 672 PutnamBench problems at the stated high test-time budget and Pass@8 setup.
  • Mistral reports 87% on FATE-H and 34% on FATE-X, plus FLTEval pass@1 of 28.9 and pass@8 of 43.2.
  • Benchmark prompts, budgets, tools, verifier policy, repository state, and comparison conditions matter. Record them before comparing systems.

Official sources

  1. Mistral's Leanstral 1.5 release — release claims, training, evaluations, license, weights, and setup.
  2. Current Mistral model card — preview status, live alias, context, output, price display, features, and weights marker.
  3. LeanstralSafeVerify — the verifier fork linked from Mistral's training description.

The release and model card disagree on the API name and active-parameter rounding. This page preserves both records instead of choosing an undocumented replacement.

Continue by decision intent

Lean 4 proof workflow Plan an auditable compiler-in-the-loop task from theorem intake through clean reproduction. Robostral Navigate Use the separate embodied-navigation record when the task is physical robot movement. Robot evaluation guide See how a single-camera navigation claim translates into a safety-gated evaluation. More model guides Browse practical access, evaluation, and implementation decision records.

Leanstral 1.5 questions, answered

Leanstral 1.5 is Mistral's specialized Lean 4 model for formal proof engineering, automated theorem proving, autoformalization, and repository-scale proof tasks.
Mistral released downloadable weights under Apache 2.0. Review the license, weight repository, dependencies, and deployment controls for the exact artifact you use.
Mistral's July 2 release names leanstral-1-5, while the current model card checked July 30 lists labs-leanstral-1-5. Query the live Mistral model catalog and test the returned identifier instead of hard-coding from this page.
The release calls the endpoint free and the checked model card displays $0 during Public Preview. Preview pricing, quotas, eligibility, and service terms can change, so verify the signed-in console before a workload.
No. Compile it with the intended Lean and dependency versions, inspect assumptions and unsafe constructs, review the changed theorem boundary, reproduce from clean state, and obtain appropriate human review.
This page does not claim a Flowith mapping. Mistral's weights and API are separate surfaces, so verify the live Flowith workspace independently.