Mistral · Theorem Proving
An AI that helps
prove theorems.
Mistral released Leanstral 1.5, a model tuned specifically to assist with automated theorem proving in Lean 4. It's the third specialist line from the company, and it makes the "specialist over generalist" strategy a real, visible bet.
The Baseline
Proving used to be
strictly a human job
Lean 4 is a language for machine-checking mathematical theorems. Building a proof means stacking tactics — small, checkable operations. Correctness is guaranteed by the machine, but the "which tactic comes next" leap has stayed on the human side.
Off-the-shelf LLMs can already list tactic candidates, but the hit rate stays too low for real work — most attempts hover at "roughly plausible." When a proof runs into a dead end, generalists rarely suggest something actually useful, and that's exactly the opening for a specialist.
Leanstral 1.5
Specialization
reads the next move
Mistral's Leanstral 1.5 is trained specifically for Lean 4 proof work.
Mistral released Leanstral 1.5, a model tuned for automated theorem proving — a real specialist for Lean 4. Unlike a generalist LLM, everything about it is optimized to judge tactic viability and reason about proof-tree branches.
Mathematicians don't get to hand off proofs whole-cloth. But the cognitive time spent finding the next move shrinks meaningfully. It proposes candidates from the current goal state, suggests different branches when the current one dead-ends — the same way a chess engine turned human thinking into a set of assisted lines rather than a replacement.
The Lineup
Specialist number
three
Mistral shipped Mistral OCR 4, a document-understanding specialist, on 6/24. General LLMs plus document intelligence plus math reasoning gives Mistral three parallel specialist lines. Where OpenAI and Anthropic bet on "one flagship or a small tier," Mistral is consolidating a distinct posture: "purpose-built by domain."
Who Feels It
A narrow model
with a narrow audience
Math researchers & PhD students
Less time spent hunting for "the next tactic" when a proof dead-ends. Directly useful for anyone touching Lean 4 projects.
Math education
Good for showing "worked examples" of proof reasoning step by step. Still needs an educator's framing — it isn't a solo tutor.
General coding use
This specialist barely touches "normal" code generation. Keep your generalist model for the day-to-day.
The Frontier
An alternative
to "one flagship"
Leanstral's real interest is that Mistral is unabashedly not competing on generality. While frontier labs race between "one flagship" and "small tier," Mistral quietly accumulates independent lines by use case. Not one big bet — a bundle of small, sharp ones.
That reshapes the mental model of "what to hand to AI." It's no longer "roll out one general model." It's "stack purpose-built ones by domain." For enterprises considering deployment, that alternative is now genuinely on the table.