Mistral released Leanstral-1.5-119B-A6B
Mistral has released Leanstral-1.5-119B-A6B, a free Apache-2.0 licensed model with 6B active parameters, showing significant performance upgrades in formal verification. It achieves state-of-the-art results on several benchmarks, including FATE-H and FATE-X.
- Leanstral 1.5 achieves state-of-the-art results on FATE-H and FATE-X benchmarks
- The model solves 587 out of 672 PutnamBench problems
- It is trained through mid-training, supervised fine-tuning, and reinforcement learning with CISPO
- Leanstral 1.5 is released under the Apache-2.0 license, making it free and open-source
The Leanstral 1.5 model is a notable upgrade, trained through a combination of mid-training, supervised fine-tuning, and reinforcement learning with CISPO. This approach enables the model to excel in agentic proof engineering.
The model's performance is demonstrated through its achievements on various benchmarks. For instance, it solves 587 out of 672 PutnamBench problems and achieves state-of-the-art results on FATE-H with 87% and FATE-X with 34%. These results underscore the model's capabilities in formal verification.
The release of Leanstral 1.5 is significant, given its free and open-source nature under the Apache-2.0 license. This makes it accessible to a wide range of developers and researchers, potentially accelerating advancements in AI and related fields.
The model's architecture and training methodology are designed to improve its performance in specific areas. By leveraging reinforcement learning with CISPO, the model can learn from its interactions and adapt to complex problem-solving tasks, particularly in formal verification and proof engineering.
Source: Mistral released Leanstral-1.5-119B-A6B. Read the full piece at the source.
offers a powerful tool for formal verification and proof engineering
advances the field of AI with significant performance upgrades
- CISPO
- a reinforcement learning framework
- PutnamBench
- a benchmark for formal verification problems
- FATE-H and FATE-X
- benchmarks for evaluating AI model performance
UN Artificial Intelligence Panel Launches Report - Havana Times
International Conference on Machine Learning (ICML) 2026 - Apple Machine Learning Research
