Mistral AI Launches Leanstral 1.5, Enhanced Open-Source Code Agent for Mathematical Proofs
Key Takeaways
- ▸Leanstral 1.5 is an updated open-source code agent model optimized for Lean 4 formal proof assistant, part of Mistral's Small 4 family
- ▸The model features 119B parameters with efficient MoE architecture (128 experts, 4 active), 256k context length, and multimodal input capabilities
- ▸Supports extended, complex tasks requiring hours of computation for mathematical proof generation and software verification
Summary
Mistral AI has released Leanstral 1.5, an updated open-source code agent model designed specifically for Lean 4, a proof assistant used for expressing complex mathematical objects and software specifications. Built as part of the Mistral Small 4 family, the model features 119B parameters with an efficient Mixture of Experts (MoE) architecture containing 128 experts with 4 active per token, offering cost-effective performance compared to closed-source alternatives.
Leanstral 1.5 includes several advanced capabilities: a 256k token context length, multimodal input support for both text and image, and tunable reasoning effort levels that users can adjust based on task complexity. The model is engineered to handle extended computational tasks requiring hours of work, making it particularly suited for complex mathematical proof generation and software verification scenarios. Its efficient architecture balances performance with cost-effectiveness, opening access to formal mathematics research for broader audiences.
The model is available through multiple deployment options: Mistral Vibe, a CLI-based agent framework for direct use, or local vLLM servers for production-ready inference pipelines. This flexibility positions Leanstral 1.5 as an accessible alternative to proprietary formal verification tools, enabling researchers and developers in mathematics and software correctness verification to leverage advanced AI capabilities.
- Deployable via Mistral Vibe CLI or local vLLM servers, providing flexible options from cloud API access to production inference
