Meituan releases LongCat 2.5 artificial intelligence model running on domestic Chinese chips

Executive Briefing
“Meituan has announced the LongCat-2.5 large language model featuring 1.6 trillion...”
Key Takeaways
- 1.LongCat-2.5 possesses a total of 1.6 trillion parameters using a mixture-of-experts architecture, where each token activates approximately 48 billion parameters during processing.
- 2.The model was trained using real tasks extracted from GitHub pull requests within a network-isolated environment.
- 3.Meituan optimized the model by compressing weights down to a 4-bit format for direct compatibility with domestic chip series such as Huawei Ascend.
Meituan has announced the LongCat-2.5 large language model featuring 1.6 trillion parameters, designed to operate entirely on domestic Chinese hardware without relying on Nvidia chips. The system releases its weights under the MIT license and supports processing text, images, and video.
Key Developments and Technical Context
LongCat-2.5 possesses a total of 1.6 trillion parameters using a mixture-of-experts architecture, where each token activates approximately 48 billion parameters during processing.
The model was trained using real tasks extracted from GitHub pull requests within a network-isolated environment.
Meituan optimized the model by compressing weights down to a 4-bit format for direct compatibility with domestic chip series such as Huawei Ascend.
Strategic Analysis and Industry Implications
The development of large-scale models optimized specifically for alternative hardware highlights efforts to reduce reliance on foreign accelerators under export restrictions.
Outlook and Key Factors to Watch
The capability benchmark results are currently self-published by Meituan and lack independent verification from external laboratories.







