Leanstral 1.5 Release Notes

Leanstral 1.5 Release Notes

Leanstral 1.5 optimizes Lean 4 formal proof engineering

Mistral AI has released Leanstral 1.5, a specialized model designed for automated theorem proving and autoformalization within the Lean 4 ecosystem. The model utilizes a Mixture-of-Experts (MoE) architecture with 119 billion total parameters, of which 6.5 billion are active during inference.

Technical Specifications and Capabilities

Leanstral 1.5 provides a high-context window and integrated tool support for formal verification workflows:

  • Model Architecture: 119B total parameters / 6.5B active parameters.
  • Context Window: 256k tokens.
  • Primary Use Cases: Automated theorem proving and autoformalization for Lean 4.
  • API Integration: The model supports Chat Completions, Function Calling, Structured Outputs, and Predicted Outputs via the Mistral API.
  • Additional Tooling: Integration with OCR, BBox Extraction, and Document Q&A is available for processing technical documentation alongside proofs.

Integration with Automated Theorem Proving (ATP) Tools

Leanstral 1.5 is designed to be used within agentic workflows for formal verification. For example, the open-source Python package OpenATP provides a CLI for agentic automated theorem provers and has integrated support for Leanstral via Mistral’s Vibe harness.

Community Feedback and Implementation Challenges

While the technical specifications of Leanstral 1.5 are targeted at a niche high-precision field, community discussion on Hacker News highlights several friction points regarding accessibility and deployment:

  • Access Issues: Users reported difficulties enabling "labs" models within the Mistral console, citing unspecified errors and a lack of responsive customer support.
  • Weight Availability: There is confusion regarding the model's weight policy. While documentation suggests Apache-licensed weights, some users noted a lack of clear download links for the 1.5 version on Hugging Face, where only earlier snapshots appear to be available.
  • Market Positioning: Some developers questioned the broader utility of specializing in Lean 4 over other proof assistants like Coq, while others noted that languages like Lean 4 and Idris 2 are particularly well-suited for LLMs due to the additional guarantees they provide during the coding process.

"Lean 4 and Idris 2 are underrated, and likely great for LLM's to code in (since they provide additional guarantees)"

Availability

Leanstral 1.5 is available as a labs model (labs-leanstral-1-5) through the Mistral AI platform. It supports a wide array of Mistral's standard API features, including FIM (Fill-In-the-Middle) completions and batch processing.

Sources