Skip to main content

Mistral AI's New Open-Source Model Makes Math Research More Accessible

Mistral AI has just dropped a new open-source model that could change how mathematicians and computer scientists work. Called Leanstral 1.5, it's built specifically for Lean4, a language used for formal mathematical proofs. And the numbers are impressive.

What Makes Leanstral 1.5 Special?

The model packs 119 billion parameters in total, but here's the clever part: it only activates 6 billion of them at any given time. That means it's powerful without being a resource hog. Licensed under Apache-2.0, it's free for anyone to use, modify, or build upon.

Crushing Benchmarks

Leanstral 1.5 aced the miniF2F formal mathematics benchmark, scoring a perfect 100% on both the validation and test sets. That's not just good—it's flawless.

But the real test was PutnamBench, a collection of notoriously difficult math competition problems. Out of 672 Lean4 problems, the model solved 587. To put that in perspective, these are problems that stump even top math students.

In abstract algebra, the model dominated the FATE benchmark series. It hit an 87% success rate on the master-level FATE-H test and 34% on the doctoral-level FATE-X test—both records for this type of model.

The Cost Factor

Here's where things get really interesting. Mistral AI says Leanstral 1.5 dramatically cuts the cost of scientific trial and error. For example, solving a PutnamBench problem with Leanstral 1.5 costs about $4 on average. Compare that to Seed-Prover1.5, which costs over $300, or Aleph Prover, which runs between $54 and $68. That's a huge difference.

This cost reduction could bring high-precision proof assistance out of the lab and into everyday research. Instead of burning through budgets, researchers can now experiment more freely.

Bug Hunting Superpowers

Leanstral 1.5 isn't just about math—it's also a sharp-eyed code reviewer. In tests across 57 code repositories, the model flagged 47 violations. Of those, 11 turned out to be real defects. Even more impressive: five of those bugs had never been reported on GitHub before. That's a strong sign that AI can help with program verification and security reviews.

What This Means

With Leanstral 1.5, the barriers to formal mathematical proof are coming down. By slashing computing and economic costs, Mistral AI is making it easier for researchers to focus on the big ideas instead of getting bogged down in tedious calculations. This model could accelerate progress in mathematics and computer science, helping more people contribute to these fields.

Image

Key Points

  • Open-source model Leanstral 1.5 is designed for Lean4 formal proofs, with 119B total parameters and 6B activated.
  • Perfect score on miniF2F benchmark; solved 587/672 PutnamBench problems.
  • Cost advantage: average $4 per problem vs. $300+ for competitors.
  • Bug finding: identified 11 real defects in code, including five previously unreported.
  • Licensed under Apache-2.0, free for all.