mistralai/Leanstral-2603 · Hugging Face
The first open-source AI agent for Lean 4 proof engineering, with 256k context and multimodal vision.
Mistral AI has released Leanstral-2603, a groundbreaking open-source code agent specifically engineered for the Lean 4 proof assistant. As part of the Mistral Small 4 family, this 119B-parameter model employs a sophisticated Mixture of Experts (MoE) architecture with 128 experts, activating only 6.5B parameters per token for remarkable computational efficiency. It boasts a substantial 256k-token context window and accepts both text and image inputs, making it a truly multimodal tool. Released under the permissive Apache 2.0 license, Leanstral-2603 is positioned as a cost-effective, high-performance alternative to proprietary proof engineering agents.
Leanstral-2603 is designed for "proof agentic" scenarios, meaning it can autonomously navigate and construct complex mathematical proofs and software specifications. Its capabilities include analyzing formal statements about objects like perfectoid spaces or verifying properties of Rust code fragments. The model is optimized for Mistral's tool-calling framework and demonstrates strong system prompt compliance, ensuring reliable and predictable behavior in automated workflows. With multilingual support across 12 languages and best-in-class inference speed, it provides a versatile foundation for researchers and developers in formal methods, theorem proving, and high-assurance software engineering.
- First open-source AI agent for Lean 4 proof assistant, enabling formal verification of complex math and software specs.
- Uses a 119B-parameter MoE architecture (128 experts, 4 active) for efficiency, activating only 6.5B parameters per token.
- Features a 256k-token context, multimodal vision, and Apache 2.0 license for commercial use, challenging closed-source alternatives.
Why It Matters
Democratizes advanced formal verification and theorem proving, allowing researchers and engineers to build high-assurance systems with open tools.