# Mistral AI Open-Sources Formal Verification LLM Leanstral 1.5

- Category: Flash (快讯)
- Published: 2026-07-06T11:48:00.000Z
- Source date: 2026-07-06
- HTML: https://www.88z.ai/en/news/flash-20260706-f4cbace19bc74232
- Markdown: https://www.88z.ai/en/news/flash-20260706-f4cbace19bc74232.md
- Other language: https://www.88z.ai/news/flash-20260706-f4cbace19bc74232.md

Mistral AI has released the open-source model Leanstral 1.5, optimized for automated theorem proving in Lean 4. Featuring a 119B parameter MoE architecture, it supports ultra-long context and multimodal inputs, achieving top performance on math benchmarks, and is now available for download and online use.

## At a glance

- Leanstral 1.5 uses a 119B MoE architecture activating 6.5B parameters per token, supporting 256k context and multimodal inputs.
- The model achieves SOTA performance on math benchmarks like miniF2F and is open-sourced on HuggingFace for local and online use.

## The update

Mistral AI has officially open-sourced Leanstral 1.5, a large language model specifically optimized for automated theorem proving in Lean 4. The model utilizes a Mixture of Experts (MoE) architecture with a total of 119 billion parameters, yet it activates only 6.5 billion parameters per token during inference, significantly improving computational efficiency. Leanstral 1.5 supports an ultra-long context window of up to 256k tokens and can process mixed text-and-image inputs, allowing it to handle complex mathematical proof tasks more flexibly. In terms of performance, the model has achieved state-of-the-art (SOTA) results on mainstream mathematical proof benchmarks such as miniF2F and PutnamBench, demonstrating strong logical reasoning and formal verification capabilities. Currently, Leanstral 1.5 is available on HuggingFace, enabling developers to either deploy it locally to ensure data privacy or experience the model's features directly through online channels, thereby lowering the barrier to entry.

## Sources

- [行业快讯](https://www.88z.ai/flash/cite/flash-20260706-f4cbace19bc74232) (2026-07-06)

This article is an AI-assisted synthesis by 88Z AI based on the public sources below. Product claims and availability are attributed to their publishers. It does not represent independent testing or first-hand reporting.