Meituan releases 1.6-trillion-parameter model trained on Chinese hardware
The new LongCat-2.5 model achieves high performance on coding benchmarks without using any Nvidia GPUs, signaling a shift in hardware independence.
The short version
- LongCat-2.5 is a 1.6-trillion-parameter mixture-of-experts model with 48 billion active parameters per token.
- The model was trained entirely on Chinese hardware, specifically using a cluster of over 50,000 chips.
- It supports multimodal inputs including text, images, and video, with a context window of up to one million tokens.
- Meituan released the weights under the MIT license, allowing for commercial use and redistribution.
- The model achieves competitive scores on coding benchmarks, such as DeepSWE, where it ranks near the top of the table.
Why it matters
This release demonstrates that high-performing AI can be developed independently of Nvidia's dominant GPU ecosystem. It provides a blueprint for how companies can build powerful, open-weight models using domestic hardware and optimization techniques.
A new standard for agents
Meituan, the Chinese food delivery company, has officially released the weights for LongCat-2.5. This model is specifically designed to power AI agents that can complete long, multi-step tasks autonomously. It is particularly well-suited for complex coding workflows where a model must navigate multiple files and logical steps.
The team at Meituan built LongCat-2.5 to compete with leading Western models. In several benchmarks, the model performs at a level similar to models like Claude Opus 5. It is designed to handle a massive context of up to one million tokens through its API while generating responses as long as 128,000 tokens.
The model also features multimodal capabilities, meaning it can process text, images, and video. Unlike some models that simply attach an image encoder to a pre-existing text model, LongCat-2.5 converts images into tokens. This allows the model to process visual data alongside text in a unified way, which helps with visual coding and 3D tasks.
Hardware independence in action
One of the most significant aspects of LongCat-2.5 is that it never touched an Nvidia GPU during its development. The model was trained entirely on Chinese hardware, specifically a cluster of more than 50,000 Chinese AI chips. This highlights a growing capability to produce high-end AI outside of the standard Nvidia ecosystem.
To make the model efficient on these chips, Meituan used a mixture-of-experts architecture. This means the model only activates a small subset of its specialized subnetworks for each token it processes. Out of the 1.6 trillion total parameters, only about 48 billion are active at any given time, which helps manage the massive scale of the model.
Meituan also optimized the model for Chinese chips by quantizing the weights to 4-bit precision. This allows the hardware to process the data natively. In load tests, this optimization increased throughput for prompt processing by 43.3 percent and for response generation by 42.77 percent compared to an 8-bit version, all while maintaining the same level of accuracy.
Training for real-world tasks
The developers used a rigorous training process to ensure the model could solve real software problems. They utilized actual software tasks from GitHub pull requests and recreated instances where previous agents had failed. To ensure the model was actually solving the problems rather than using shortcuts, the team blocked internet access during training.
The training environment was stripped of Git history and cached files to prevent the model from
Deployment and performance metrics
For those interested in self-hosting the model, Meituan provided a specific guide for a cluster of 194 Huawei Ascend accelerators from the Atlas A2 series. This setup allocates two units for image processing, 64 for prompt ingestion, and 128 for response generation. Notably, the documentation does not provide any instructions for using Nvidia GPUs.
The model's performance varies depending on the difficulty of the task. On the harder variant of the Terminal Bench, LongCat-2.5 scored 38.5 points. This puts it slightly ahead of GPT-5.6 Sol, which scored 37.3, but it remains significantly behind Opus 5, which reached 53.9 points.
Access and community resources
Users who want to test the model's capabilities can try it out through a chat interface at the official website. The model is also available via an API that supports standard formats used by OpenAI and Anthropic. This allows it to plug into existing agent harnesses like Codex, Hermes, and Claude Code.
The weights are shared under the MIT license, which permits commercial use and redistribution. For more updates on the AI landscape, readers can subscribe to The Decoder for a weekly newsletter. This subscription includes ad-free reading, a weekly news recap, and access to deep-dive frontier reports six times a year.
Sources and links
- Codeself-hosting guide github.com
- Modelweights are available huggingface.co
- Modelseparate repository huggingface.co
- ReportingThe Decoder: the article we sourced
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.
Listen to the AI Wire headlines
Listen to the AI WireListen to our daily AI shows every morning: on your commute, in the shower, wherever you are
Listen to the daily showsFirst reported by The Decoder · Read the original. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
