Mistral's Lean 4 formal-proof model for automated theorem proving and autoformalization.
Last verified · 2026-07-01 · by Moe Ameen
Leanstral 1.5 is a specialized model from Mistral, the French AI lab, built for formal proof engineering in Lean 4 rather than general writing or chat. Mistral released it on June 30, 2026 as an update to the original Leanstral (March 2026), optimised for two jobs: automated theorem proving and autoformalization. Despite sitting in a chat-style API, it is not a general text-generation model, and treating it as one misses the point of what it does.
The two jobs are worth pinning down. In theorem proving, you hand it a proof goal and it proposes the Lean 4 steps, which the Lean 4 proof assistant then mechanically checks — so the output is verified correct, not just plausible-looking. In autoformalization, it converts a human-readable mathematical statement or a software specification into Lean 4-compatible definitions and theorems. That combination matters most where a single bug is expensive: cryptography, aerospace, financial modeling, and safety-critical code.
Under the hood it is a Mixture-of-Experts model — Mistral lists 119B total parameters with 6.5B active per token — paired with a 256k-token context window large enough to hold long proof files. The MoE design is where "efficient" comes from: it runs closer to the cost of a small model while drawing on a much larger parameter pool. Mistral said the original Leanstral beat Claude Sonnet by 8 points at pass@16 while costing about 15x less; it has not published new head-to-head benchmarks for 1.5, so treat the March figures as the reference point rather than a fresh claim.
Leanstral 1.5 is listed at $0 on Mistral's Labs tier and can be tried in the console playground as `labs-leanstral-1-5`. The original Leanstral shipped open-weights under Apache 2.0 on Hugging Face; at launch the 1.5 page confirms only the free Labs access, so if open weights or self-hosting matter to you, verify their status for 1.5 specifically before building on it.
Be honest about the overlap first: there is almost none. Leanstral 1.5 proves theorems and verifies code; Kompozy generates and publishes content. A raw Lean 4 proof is not a social post, and Kompozy will never write you a proof. The people who pair the two are technical educators, dev-tool founders, and researchers who ship formal-methods work and then need an audience for it. You do the proof in Leanstral; the story of that proof is what goes into Kompozy. Write a short plain-language summary — "we formally verified our payment engine in Lean 4, here's how" — drop it into Kompozy as a source, and it fans that one result into a carousel that walks the proof sketch, a LinkedIn post and an X thread in your own voice via the Persona Brief, a blog article, and a short persona or avatar video explaining the outcome, all rendered on-brand through HyperFrames and scheduled across the nine connected platforms from one queue.
The gap this closes is specific. Formal-methods work is genuinely hard to market: the audience that cares is small, technical, and scattered, and most researchers dread the "now go tell people" half of shipping. Leanstral gives you something that is actually worth posting — a verified result, not marketing fluff — and Kompozy is the layer that turns your one paragraph about it into a week of technical content without you opening six apps. Do not expect a mechanical pipe from proof to post; the handoff is a human write-up of what you proved. That write-up is where Kompozy earns its keep.
Leanstral 1.5 is a specialized model from Mistral, released June 30, 2026, built for formal proof engineering in Lean 4. It is optimised for automated theorem proving (proposing Lean 4 proof steps that the Lean checker verifies) and autoformalization (turning human-readable math or specs into Lean 4 definitions and theorems). It is a Mixture-of-Experts model with 119B total parameters, 6.5B active per token, and a 256k-token context window.
No. Although it is served through a chat-style API, Leanstral 1.5 is tuned for Lean 4 proof work — theorem proving and autoformalization — not general writing, summarizing, or conversation. Use a general model for those tasks and Leanstral for verified proofs.
Mistral lists Leanstral 1.5 at $0 on its Labs tier, available to try free in the console playground as labs-leanstral-1-5. Labs pricing and availability can change, so confirm current terms in Mistral's console before you build on it.
Leanstral 1.5 (June 30, 2026) iterates on the original Leanstral released in March 2026. Mistral has not published new head-to-head benchmarks for 1.5; the original was reported to beat Claude Sonnet by 8 points at pass@16 while costing about 15x less. Treat those figures as the prior release's numbers, not confirmed gains for 1.5.
No. Leanstral produces proofs and formalized statements, not content or media. To turn a verified result into posts, carousels, a blog, a newsletter, or a persona video, write a short plain-language summary and pass it to a content engine like Kompozy, which generates each format in your voice and publishes across platforms.