Meituan is putting a very specific hardware claim at the center of its new AI model: LongCat-2.0, the company says, was trained end-to-end on domestically developed chips.
That matters more than the model’s headline size. LongCat-2.0 is described by Meituan as a 1.6-trillion-parameter model with a one-million-token context window. The company also says the model’s performance is comparable to Google’s Gemini 3.1 Pro. Those are big claims on their own, but the more politically and commercially important one is about the compute stack underneath it.
Meituan says LongCat-2.0 completed both training and inference on a 50,000-chip domestic compute cluster. That training-hardware claim has not been independently verified, and outside observers have limited visibility into the company’s infrastructure. Still, the announcement is clearly framed around one of the central questions in AI right now: how far China’s large-model ecosystem can go without relying on the most advanced US-controlled chips.
Why the chip claim matters
Running an AI model and training one are not the same problem. Inference is the work a model does after training, when it answers prompts or powers an application. Pre-training is the heavy lift, where a model learns broad patterns from large datasets and burns through enormous amounts of compute.
That distinction is why Meituan’s wording is important. Plenty of companies can talk about serving AI models on alternative hardware. A claim of end-to-end domestic training, if borne out, points to something more substantial: a large model built and operated without leaning on the foreign accelerators that have dominated frontier AI.
The backdrop is the US effort to restrict China’s access to advanced AI chips on national-security grounds. Beijing, in turn, has been pushing domestic semiconductor development and encouraging Chinese technology companies to reduce dependence on foreign hardware. LongCat-2.0 fits into that broader push, but the details that would let outsiders fully validate the training setup remain limited.
What Meituan says LongCat-2.0 includes
Meituan describes LongCat-2.0 as a trillion-parameter-class model with a very large context window. It has also positioned the system as open-source, which could allow developers and researchers to test the model’s behavior, performance, and deployment requirements more directly.
The open-source piece is commercially important. For developers, model weights are the difference between reading a benchmark claim and actually seeing how a model behaves in practice. For Meituan, opening the model can help drive adoption, attract technical scrutiny, and make LongCat-2.0 part of the broader conversation around Chinese AI infrastructure.
The strongest claims, however, still need to be separated from what can be easily tested. Developers can evaluate the model’s output quality, context handling, latency, and resource needs once they have access to the relevant release. They cannot easily confirm, from the model alone, exactly what chips were used across the full training run.
| Claim | Why it matters | What remains unclear |
|---|---|---|
| 1.6 trillion parameters | Positions LongCat-2.0 as a very large model | Parameter count alone does not prove quality or efficiency |
| One-million-token context window | Suggests support for very long prompts and documents | Real-world usefulness depends on accuracy across long contexts |
| Domestic end-to-end training and inference | Would signal progress in China’s AI compute stack | The hardware claim is difficult to verify externally |
| Open-source release | Lets developers inspect and test the model more directly | Public testing can validate performance, not the full training supply chain |
Why Meituan is an interesting company to watch
Meituan is better known for food delivery and local services than for frontier AI. That makes LongCat-2.0 a useful signal in a different way: large AI models are no longer just prestige projects for specialist labs. They are becoming infrastructure for internet platforms that depend on logistics, search, recommendations, customer service, routing, and demand forecasting.
For a company operating at Meituan’s scale, cheaper and more controllable AI infrastructure would have practical value. Domestic compute could reduce exposure to export-control changes and make long-term capacity planning less dependent on foreign chip availability. That does not prove LongCat-2.0 delivers those advantages today, but it explains why Meituan would want to make the hardware story central to the release.
The company’s move also reflects a wider pattern among Chinese technology giants. AI model development is increasingly treated as a strategic layer of the business, not a side experiment. The companies that can train, tune, and deploy models on infrastructure they can reliably access will have more room to iterate, whether the use case is consumer search, enterprise tools, logistics, or developer platforms.
What buyers and developers should take from it
For enterprise teams, LongCat-2.0 is not simply a model-size story. The relevant question is whether it can perform reliably on the workloads buyers actually care about, and whether its deployment requirements make economic sense compared with better-known alternatives.
The open-source release, if accessible and complete enough for meaningful testing, gives technical teams a clearer path to evaluation. They can benchmark it against their own documents, prompts, latency targets, and infrastructure constraints instead of relying only on headline comparisons.
For chip buyers, cloud providers, and AI infrastructure teams, the bigger signal is competitive pressure. If large Chinese models can be trained and served on domestic clusters, even with caveats, that would make the AI hardware market less dependent on a single supply chain. If the claim proves overstated, it still shows where Chinese platforms want the market to believe things are heading.
The caveat is the story
The most important thing to hold onto is the difference between a testable model and a testable infrastructure claim. LongCat-2.0 can be evaluated by developers once they can run it or inspect its published materials. The assertion that it was trained end-to-end on domestic chips is much harder to prove from the outside.
That does not make the claim irrelevant. It makes it consequential and unresolved. Meituan is presenting LongCat-2.0 as evidence that China’s AI stack can move up the value chain despite US chip restrictions. The market response will depend on whether the model performs well in public testing, and whether the domestic-compute story gains more concrete validation over time.
For now, LongCat-2.0 is best understood as both a model release and a strategic message. Meituan is not only competing for developer attention. It is also arguing that the hardware bottleneck around large AI models may be less fixed than Washington intended.