Mistral — Leanstral 1.5
Read the code
as if it were a proof.
Mistral's new open-source model Leanstral 1.5, released on July 5, tops formal-math benchmarks and finds real bugs in production code at the same time. "Audit with a mind that can prove theorems" — a quiet experiment has begun.
A Second Specialist
Mistral gets its second dedicated specialist
Following June 24's OCR 4, the second task-focused model has arrived — this one tuned for math and code.
Until recently the Mistral lineup was all generalists, built around the Medium and Large tiers. The turning point was June 24's document-analysis OCR 4. And on July 5, Mistral released a second task-specialist, the open-source Leanstral 1.5. The name is deliberate — a nod to the Lean proof assistant crossed with the Mistral core.
The license is commercial-friendly Apache 2.0. Weights are on Hugging Face; you can run it standalone or drop it into an existing inference stack. It is smaller than the flagship generalists, but its pitch is that inside its narrow domain it lands close to the frontier.
Two Muscles
Formal proofs and real-world bugs, from the same model
Not "great on the leaderboard, useless in practice" — nor the opposite. It sits between them.
The first muscle is formal proof. On the main Lean 4 benchmarks, Leanstral 1.5 lands where smaller models could not reach — solid at the "detect the broken step and rewrite it" side of proof work, not just at chaining tactics.
The second is picking real defects out of production code. Mistral's technical write-up walks through cases where Leanstral 1.5 was pointed at existing OSS repositories and surfaced concrete bugs that had slipped past review. Proof work and static-analysis-style work travel between the same model.
Where It Fits
Where it actually helps
Code audit
Work previously handed to external security firms or SAST tools — the "shortlist of suspicious spots" step — now has an open-weight option. Because it is Apache 2.0 you can run it against code that must never leave the building.
Math startups and researchers
Research-oriented Lean 4 work gains a commercial-friendly open-weight assistant. Teams that want to fine-tune a proof-oriented model and keep the result in-house get a solid starting base.
Everyday chat users
Not for you, and Mistral doesn't pretend otherwise. Conversation, summarisation and writing stay with the generalist models. Think of Leanstral 1.5 as one more specialised tool in the drawer.
Reading code with a mind that can close a proof
is a different kind of scrutiny from a review.