Meituan's free AI model nears US leaders and runs on Chinese chips
China's food delivery giant has released a large AI model for long, multi-step work that anyone can download. Meituan tuned it for Chinese AI chips, and its guide to running it uses no Nvidia hardware.

The short version
- Meituan, China's food delivery giant, has released LongCat-2.5, a large AI model that anyone can download and use for free.
- It is built for long, multi-step work like writing software, and comes close to leading US models on several of Meituan's tests.
- Meituan says key parts of its training ran on Chinese chips. It also published a guide to running it on 194 Huawei Ascend chips, with no Nvidia hardware involved.
What Meituan released
Meituan's LongCat team announced LongCat-2.5 on 1 November 2026, then published the model on Hugging Face. It is released under the MIT license, which allows commercial use.
It is a very large model, at 1.6 trillion internal settings. But it uses only a small part of itself for each word it reads, so it costs less to run than its size suggests.
It is designed for long jobs that take many steps, such as building a whole app or fixing bugs across a large codebase. It also understands images and video. Meituan's demos range from a business operations analysis and a feedback report to 3D product showcases and an event invitation tool.
How good it is
Meituan compares LongCat-2.5 with leading US models on more than two dozen tests. On a widely used test of working in a computer's command line, it scored 88.4. Meituan lists Anthropic's top model, Claude Opus 5, at 89.1 on the same test. On a harder version of that test, it trails Opus 5 by a wide margin.
Meituan ran many of these tests itself and says its own scores are indicative only. On some tests, other models in the table score higher.
To keep the model honest during training, Meituan blocked internet access in its practice tasks and removed old code history. It also had a separate AI check the work for shortcuts.
Why the chips matter
Meituan says key parts of its training ran on what it calls domestic AI accelerators, meaning Chinese-made AI chips. It compressed the model so those chips can run it efficiently. Meituan says accuracy stayed comparable to the uncompressed model.
It also published a step-by-step guide for running the model on a cluster of 194 Huawei Ascend chips. The guide gives no route for Nvidia hardware.
A capable model tuned for Chinese chips makes that hardware a more practical choice, even outside China.
What it means for companies in Indonesia
Free, downloadable models now come from more places. That gives you more bargaining power with AI suppliers. It also lets you run AI on your own servers when data can't leave the building.
But a model's origin now matters for more than performance. Some buyers, especially banks and government suppliers, will ask where a model came from and what hardware it needs. Questions about export controls and security reviews can slow a project down.
Treat free models like any other supplier. Test them on your own work and in your own languages. Have your security team review them before use.
What to do or watch
- Keep a short list of free models you could switch to if a paid supplier raises prices.
- Test any candidate on real Indonesian-language tasks, not just the maker's scores.
- Ask your security and legal teams for a clear policy on models from different countries.
- Watch whether Indonesian cloud providers start offering non-Nvidia hardware for running AI.
