- Published on
Proof Abundance for All
- Authors
- Name
- DeepHorus
Leanstral 1.5, a new open-source model for formal verification in Lean 4, has been released by the Leanstral team, achieving state-of-the-art results across multiple benchmarks while remaining free and accessible. The model, licensed under Apache-2.0, features 119 billion total parameters but only 6 billion active parameters, making it computationally efficient. It is designed to make rigorous formal methods practical for real-world software verification and mathematical proof engineering.
Formal verification tools have traditionally been expensive and difficult to use, limiting their adoption beyond academic settings. Leanstral 1.5 aims to change that by offering a powerful, open-source solution that can prove complex theorems and verify code properties. The model builds on the original Leanstral release and has been refined through a three-stage training process: mid-training, supervised fine-tuning, and reinforcement learning with a novel algorithm called CISPO.
On standardized benchmarks, Leanstral 1.5 saturates the miniF2F dataset, achieving 100% on both validation and test sets. It solves 587 out of 672 problems on the PutnamBench, surpassing previous models like Seed-Prover 1.5 by 7 problems. On the advanced FATE-H and FATE-X benchmarks, it achieves new state-of-the-art scores of 87% and 34%, respectively. These results demonstrate its ability to handle everything from elementary math to graduate-level abstract algebra.
A standout feature of Leanstral 1.5 is its cost efficiency. While competing models like Seed-Prover 1.5 cost an estimated $300 per problem in computational resources, Leanstral 1.5 operates at roughly $4 per problem. This dramatic reduction in cost is achieved through its efficient architecture and training methodology, which includes multiturn interaction with the Lean compiler and a code agent environment that simulates real-world development workflows.
Beyond benchmarks, Leanstral 1.5 has proven effective in practical code verification. It was tested on 57 open-source repositories and uncovered 5 previously unknown bugs. This capability highlights the model's potential to assist human developers in finding subtle errors in complex software, bridging the gap between formal methods and everyday programming.
The release has been met with enthusiasm from the formal verification community. Researchers note that Leanstral 1.5 democratizes access to advanced proof engineering, offering a free API and full model weights on Hugging Face. By lowering barriers to entry, it enables more developers and mathematicians to incorporate formal verification into their workflows.
Looking ahead, the Leanstral team plans to continue refining the model, focusing on expanding its capabilities to larger codebases and more complex mathematical theories. With its combination of performance, cost efficiency, and openness, Leanstral 1.5 represents a significant step toward making formal verification a standard tool in software development and mathematical research.
