Chinese technology company Meituan has released the model weights for LongCat-2.5, a 1.6-trillion-parameter artificial intelligence system designed to handle complex tasks through autonomous AI agents. The model is attracting attention because Meituan says its development and deployment rely on Chinese-made semiconductor hardware rather than Nvidia GPUs, highlighting China’s efforts to build advanced AI systems without depending on American chip suppliers.
The release became available on November 15, 2026, approximately one week after the model’s initial unveiling. LongCat-2.5 uses a mixture-of-experts architecture, activating approximately 48 billion parameters for each token despite having 1.6 trillion parameters in total. Its capabilities include coding, software engineering, image and video understanding, and multi-step agent workflows. Meituan has also released the model under the permissive MIT licence, allowing developers to download, modify and commercially use its weights.
Meituan Releases LongCat-2.5 With 1.6 Trillion Parameters
LongCat-2.5 was developed by Meituan’s LongCat AI research team, which is working to establish the company as a significant player in China’s increasingly competitive AI industry.
The model is designed primarily for agent-based applications. Unlike conventional chatbots that mainly respond to questions, AI agents can plan and execute multiple steps to complete tasks such as writing software, navigating development environments and analysing information.
Its architecture contains 1.6 trillion parameters, but it does not activate all of them for every request.
Instead, LongCat-2.5 uses a Mixture-of-Experts (MoE) architecture, in which specialised groups of neural-network parameters handle different parts of a task. Approximately 48 billion parameters are active for each token.
This approach allows the model to maintain a large overall capacity while limiting the amount of computation required for individual operations.
The release also includes downloadable model weights, making it possible for developers and organisations to operate the system on their own infrastructure rather than depending exclusively on Meituan’s hosted services.
The Model Was Developed Without Nvidia GPUs
One of the most significant aspects of LongCat-2.5 is its reliance on China’s domestic semiconductor ecosystem.
According to Meituan and the model’s technical documentation, the system was developed using Chinese-made AI accelerators rather than Nvidia GPUs.
The company’s approach is particularly relevant because US export restrictions have limited Chinese technology companies’ access to some advanced Nvidia processors.
Meituan’s predecessor, LongCat-2.0, was trained on a cluster containing more than 50,000 Chinese-made AI chips, according to the company. The newer LongCat-2.5 continues the strategy of optimising AI development for domestic hardware.
The distinction matters because training a trillion-parameter model requires enormous computing resources. Historically, Nvidia GPUs have played a central role in the development of large AI systems around the world.
Meituan’s work demonstrates that Chinese companies are developing alternative hardware and software combinations capable of supporting large-scale AI development.
However, the achievement should not be interpreted as proof that Chinese accelerators have completely replaced Nvidia across the global AI industry. It demonstrates that Meituan has built and released this particular model using a domestic hardware ecosystem.
LongCat-2.5 Can Run on 194 Huawei AI Accelerators
Meituan’s deployment documentation provides another important detail about the hardware behind the model.
The company describes a configuration using 194 Huawei Ascend accelerators from its Atlas A2 series to run LongCat-2.5.
The hardware is divided into different groups according to their workloads.
| Hardware allocation | Number of accelerators | Function |
|---|---|---|
| Image processing | 2 | Handles visual input processing |
| Prompt processing | 64 | Processes incoming requests |
| Response generation | 128 | Generates model responses |
| Total | 194 | Complete deployment configuration |
This configuration illustrates how large AI models can be distributed across specialised computing resources.
The deployment guide does not provide instructions for running this version on Nvidia GPUs. Instead, it focuses on Huawei’s domestic accelerator infrastructure.
For Chinese AI developers, this is important because the availability of a complete deployment pathway can be as valuable as the model weights themselves. Developers need compatible hardware, software frameworks and optimisation tools to operate large models efficiently.
Get the day’s top stories in your inbox
One concise email. No spam, unsubscribe anytime.



