Slide 1
Mistral just open-sourced a proof model that fires six and a half billion parameters per token, and out-proves models many times its size. The weights are already public, under Apache two.
Top open prover on PutnamBench: 587 of 672
On PutnamBench, six hundred and seventy-two Putnam problems in Lean four, it solves five hundred eighty-seven. Mistral says that's the top open prover, seven ahead of Seed-Prover, and second only to one closed system.
It doesn't just solve problems. It operates the tools.
The novelty isn't the score, it's how it gets there. Only four of its hundred and twenty-eight experts fire per token, and it's trained as an agent: it drives a filesystem, a shell, and the Lean compiler, refining each proof against real feedback until it checks. That's what lets it leave the benchmark and read live code.
Beats Opus 4.6 on formal proof: pass@8 43.2 vs 39.6
Then there's this. On Mistral's FLTEval benchmark, the tiny open model scores forty-three point two at pass eight. That beats Claude Opus four point six, at thirty-nine point six, on formal proof, for about a seventh of the cost. And FLTEval is built from real pull requests in the Fermat's Last Theorem project, not a synthetic set.
5 bugs nobody had reported, in shipping code
// datrs/varinteger · src/lib.rs fn sign(value: u64) -> i64 { if value & 1 != 0 { -(((value + 1) / 2) as i64) } else { (value / 2) as i64 } }
u64::MAX is odd, so value + 1 overflows:
panic in debug, silently wrong in release. Flagged by Leanstral 1.5.
Because it can run code, Mistral pointed it at fifty-seven real repositories. It flagged forty-seven violated properties, eleven genuine bugs, and five nobody had reported before. One was a sign function in a Rust varint library that overflows on the largest unsigned integer, panicking in debug and silently returning the wrong value in release. A proof found it, in shipping code.
Read the fine print: whose benchmarks, and at what budget
Now the honest asterisks. Most of those wins are on Mistral's own evals, and the margins are thin: eighty-seven to eighty, thirty-four to thirty-three. The one independent board is PutnamBench, where it's second, behind a closed system that solves six hundred sixty-eight. And that headline five eighty-seven rides a four-million-token budget per problem. The cheap four-dollar figure sits lower on the same curve.
Open weights on Hugging Face · Apache-2.0 · free API
And it's genuinely open. Full weights on Hugging Face under Apache two, plus a free hosted endpoint if you'd rather just call it.
Slide 8
The shift is simple. Machine-checked proof used to be closed and expensive. One release made it cheap and open. Now watch whether the public Putnam leaderboard confirms these numbers out in the open.
