Skip to main content

Mistral AI's Math Tool: 119B Parameters, Only 6B Active, Cost 1% of Rivals

European AI company Mistral AI has released Leanstral 1.5, a model built specifically for formal mathematical proofs using the Lean4 programming language. The model boasts a total of 119 billion parameters, but here's the kicker: during actual reasoning, it only activates 6 billion of them. That's like having a massive library but only pulling out the exact books you need. And it's all open-sourced under the Apache-2.0 license.

Image

On core benchmarks, Leanstral 1.5 delivered almost flawless results. It scored a perfect 100% on both the validation and test sets of the miniF2F formal mathematics benchmark. In the PutnamBench math competition, it solved 587 out of 672 Lean4 problems. For the FATE series benchmarks in abstract algebra, it hit 87% at the master's level (FATE-H) and 34% at the doctoral level (FATE-X)—both best-in-class.

Cost: Just 1% of Competitors

But the real jaw-dropper is the cost. On PutnamBench, Leanstral 1.5 averages only $4 per problem. Compare that to ByteDance's Seed-Prover 1.5, which costs over $300 per problem, or Aleph Prover at $54–$68. That's a 99% reduction in reasoning costs, making formal mathematical proofs economically viable at scale.

Real-World Impact

Leanstral 1.5 isn't just a lab experiment. In practical tests, it scanned 57 code repositories and identified 47 violating attributes. Among those, 11 pointed to real code defects, and 5 were completely new issues never reported on GitHub. From math competitions to software verification, this model proves that parameter size isn't everything—efficient activation is what brings AI reasoning to the real world.

Key Points

  • Model: Leanstral 1.5, specialized for Lean4 formal proofs.
  • Parameters: 119B total, only 6B activated during inference.
  • Performance: 100% on miniF2F, 587/672 on PutnamBench, top scores on FATE benchmarks.
  • Cost: ~$4 per problem vs. $300+ for competitors.
  • Open Source: Released under Apache-2.0.
  • Real-World Use: Found 11 real code defects, 5 previously unknown.