Meituan open-sources LongCat-2.5, a 1.6T-parameter agent model trained with no Nvidia GPUs
Meituan has open-sourced the weights of LongCat-2.5, a 1.6-trillion-parameter agent model trained entirely on Chinese accelerators — with no Nvidia GPUs involved at all.

The News
The Decoder reported on November 15 that Meituan's LongCat team has open-sourced the weights of LongCat-2.5 under the MIT licence, a permissive licence that lets anyone download, use and build on the model. LongCat-2.5 is a 1.6-trillion-parameter mixture-of-experts agent model, with about 48 billion parameters active for each token it processes. An agent model is one built to carry out multi-step tasks on its own — using tools, running code and finishing jobs — rather than simply answering a single question.
The headline is the hardware. LongCat-2.5 was trained entirely on Chinese accelerators, and Meituan's self-hosting guide documents a 194-chip Huawei Ascend Atlas A2 cluster, with no Nvidia GPU guidance at all. That matters because Nvidia's chips have dominated AI training, and Chinese companies face restrictions on buying the most advanced ones. Training a model of this size with no Nvidia hardware in the picture shows that a domestic alternative can work at scale.
On performance, Meituan reports a score of 88.4 on Terminal Bench v2.1, a test of how well AI agents handle real computer tasks, just behind Claude Opus 5's 89.1. But those are vendor-reported scores with no independent reproduction yet, so they should be read as Meituan's own claim rather than a verified result. One date point to keep straight: the model was unveiled on November 2 and Meituan announced the open-sourcing on November 10; November 15 is the date of the Decoder report and the Huawei-only deployment detail. Meituan's shares closed at HK$71.55 on Monday, November 16, up 2.51%.
LongCat-2.5 was trained entirely on Chinese accelerators — a 194-chip Huawei cluster with no Nvidia GPUs at all.
Why It Matters
The hardware story is the one that will travel furthest. For years, training a frontier AI model has effectively meant training on Nvidia GPUs, and the most advanced of those chips are restricted for sale to Chinese companies. A 1.6-trillion-parameter model trained entirely on Huawei accelerators suggests that the restriction, while painful, is not an absolute barrier. That is a development with consequences far beyond one model or one company.
Open-sourcing the weights sharpens the point. By releasing LongCat-2.5 under the MIT licence, Meituan is inviting the world to download the model, run it and check its claims. That openness works in Meituan's favour if the model performs as advertised, because independent users become its best validators. It also spreads the technology widely, which is a strategic choice as much as a technical one: open models tend to attract developers, integrations and improvements from outside the company that built them.
For the wider market, the read-through is mixed. If domestic Chinese hardware can train frontier models, the leverage of chip export controls weakens, and Nvidia's grip on the training market faces a new kind of challenge — not from a rival chip that is faster, but from a whole alternative stack that is good enough. At the same time, one model does not prove an entire ecosystem. The questions that remain — cost, reliability, software support, and whether others can repeat the feat — will decide whether this is a milestone or a one-off.
Analyst's Take
Our call is Mixed. Training a 1.6-trillion-parameter model entirely on Chinese accelerators is a genuine technical achievement and a meaningful answer to the hardware restrictions facing Chinese AI companies. The open-sourcing under the MIT licence adds to that: anyone can now download the weights and test the model for themselves.
The caution is just as real. The 88.4 benchmark score is vendor-reported, with no independent reproduction yet, so the claim that LongCat-2.5 sits just behind Claude Opus 5 is still unproven. A score on one agent benchmark also says little about how the model behaves on the messy, varied work real users give it. And one model does not by itself prove that a whole domestic hardware stack can match Nvidia across the industry.
What is clear is that the hardware map of AI is changing. If Chinese teams can train frontier-scale models on domestic chips and then give the weights away, the competitive picture for chipmakers and model builders alike gets more complicated. The next test is independent: let outside labs reproduce the scores and run the model on their own tasks. Reader poll: Bullish or bearish on 3690.HK? Vote: Bullish, Bearish, Mixed, Wait for proof.
Are you bullish or bearish on this story?
By The Numbers
FAQ
What does open-sourced weights mean?
It means Meituan has released the trained model files under the MIT licence, so anyone can download them and run or study the model themselves.
What is an agent model?
It is an AI model built to carry out multi-step tasks on its own — using tools, running code and completing jobs — rather than just answering a single question.
Why does training without Nvidia GPUs matter?
Nvidia's chips have dominated AI training, and access to them is restricted for Chinese companies, so training a top model on domestic hardware shows an alternative path is possible.
👁
Source: The Decoder