Mistral AI Releases Leanstral 1.5: An Apache-2.0 Lean 4 Code Agent Model Solving 587 of 672 PutnamBench Problems
Mistral AI launched Leanstral 1.5, an Apache-2.0 licensed code agent for Lean 4 that solves 587 of 672 PutnamBench problems. The model uses a 119B MoE architecture with 6.5B active parameters.
- Leanstral 1.5 solves 587 of 672 PutnamBench problems, achieving an 87% success rate in formal theorem proving.
- The model uses a 119B parameter MoE architecture with only 6.5B active parameters per token for efficiency.
- It saturates performance on the miniF2F benchmark, indicating state-of-the-art capabilities in automated reasoning.
- Includes real bug-finding case studies and deployment code under the Apache-2.0 license.
Mistral AI has released Leanstral 1.5, a free and open-source code agent model designed specifically for Lean 4, a formal proof assistant used in mathematical theorem proving. The model achieves state-of-the-art performance by solving 587 out of 672 problems in the PutnamBench benchmark, representing an 87% success rate. This positions Leanstral 1.5 as a significant advancement in automated theorem proving and formal verification tools.
The architecture leverages a 119 billion parameter mixture-of-experts (MoE) design, where only 6.5 billion parameters are activated per token. This sparse activation approach enables high performance while maintaining computational efficiency. The model has been validated on the miniF2F benchmark, where it saturates performance, indicating it has reached or surpassed human-level capabilities in certain formal reasoning tasks.
Beyond benchmark performance, Mistral AI has included real-world case studies demonstrating the model's ability to identify and fix bugs in Lean 4 codebases. The release also comes with deployment code and documentation, making it accessible for researchers and developers to integrate into their workflows. The Apache-2.0 license ensures the model can be freely used, modified, and distributed.
Source: Mistral AI Releases Leanstral 1.5: An Apache-2.0 Lean 4 Code Agent Model Solving 587 of 672 PutnamBench Problems. Read the full piece at the source.
Provides a powerful, open-source tool for formal verification and theorem proving in Lean 4.
Enables automation of formal reasoning tasks, reducing manual verification costs in critical systems.
Offers an accessible way to learn and experiment with formal mathematics and automated theorem proving.
Demonstrates rapid progress in AI-driven formal reasoning and its potential applications in mathematics and computer science.
- PutnamBench
- A benchmark dataset of challenging mathematical problems used to evaluate formal theorem proving models.
- miniF2F
- A standardized benchmark for evaluating formal reasoning models across multiple theorem-proving languages.
- Mixture-of-Experts (MoE)
- A neural network architecture where only a subset of parameters (experts) are activated for each input, improving efficiency.
![Training transformers where every layer W = V·Uᵀ from initialization reveals a corpus-determined optimal rank - looking for arXiv endorser (cs.LG) [D]](https://images.weserv.nl/?url=external-preview.redd.it%2FQfw5SuGCt2d45VbzHurInHB_fbCrPRWPZr4XzFenJcc.png%3Fwidth%3D140%26height%3D70%26auto%3Dwebp%26s%3D6e9379fe0f90d43518578b30abf4563219025786&w=520&fit=cover&q=70&output=webp&dpr=2&we=1&il=1)