Meituan’s LongCat-2.0 is a major signal in the race to train frontier-scale AI models without relying on Nvidia hardware. The Chinese company says the 1.6 trillion parameter model was trained entirely on domestic computing infrastructure, using more than 50,000 domestically made AI ASICs and over 35 trillion tokens.
That claim matters because it ties model performance, chip supply, and national AI capacity into one result. LongCat-2.0 is not just another large model announcement. It is being presented as evidence that China can train massive AI systems on homegrown hardware even under US export controls in place since 2022.
What Meituan says LongCat-2.0 proves
Meituan framed the achievement around domestic AI infrastructure. The company said:
"LongCat-2.0 has demonstrated that we now have the capability to train large-scale models on domestic computing clusters,"
The core facts are direct: LongCat-2.0 has 1.6 trillion parameters, was trained on a cluster of more than 50,000 domestically made AI ASICs, and covered over 35 trillion tokens. Meituan did not identify the specific chip maker behind the hardware.
The timeline is also notable. The LongCat team has only existed since 2023, and its first model shipped late last year. Within that short window, the team has moved from an initial release to a trillion-parameter system positioned against leading Western models.
For readers tracking AI infrastructure, the hardware claim may be as important as the model itself. Training a model at this scale requires coordination across chips, networking, software, and data pipelines. Meituan’s announcement says that a domestic computing cluster was able to support that work from end to end.
Where the model looks competitive
LongCat-2.0’s reported benchmark results are mixed, but several are strong enough to draw attention. On some coding-focused tests, the model outperforms major Western systems named in the source article.
On SWE-bench Pro (59.5) and SWE-bench Multilingual (77.3), LongCat-2.0 tops Gemini 3.1 Pro and GPT-5.5. It does not lead every comparison, however. On those same tests, it still falls short of Claude Opus 4.7 and 4.8.
That split is important. It suggests LongCat-2.0 should not be treated as a blanket winner across model evaluation. Instead, the available figures show a model that is highly competitive in some benchmark areas while still behind other frontier systems in others.
The source also lists tests where LongCat-2.0 trails Gemini and GPT-5.5, in some cases by a wide margin. These include IFEval (90.0), IMO-AnswerBench (81.8), and GPQA-diamond (88.9). Those results create a more measured picture: LongCat-2.0 appears powerful, but its strengths are uneven across the reported benchmark set.
Why the Nvidia angle matters
The LongCat-2.0 announcement lands in a wider debate about whether Chinese AI labs can keep pace without access to the most desired Western AI chips. The source article makes the geopolitical signal explicit: despite US export controls in place since 2022, China appears to have produced its first competitive trillion-parameter model trained entirely on domestic hardware.
That does not mean every unanswered question is resolved. Meituan did not name the specific chip maker, so outsiders cannot fully assess the hardware stack from the information provided. The model also is not yet available on HuggingFace, which makes independent verification difficult.
Still, the reported training run changes the discussion. If the details hold up, the key point is not only that LongCat-2.0 exists. It is that Meituan says the model was trained without Nvidia, at massive scale, using Chinese AI chips in a domestic computing cluster.
For Washington, the message is hard to miss. Export controls were designed to limit access to advanced AI hardware. LongCat-2.0 suggests that China’s domestic chip and model-building ecosystem may be able to keep producing competitive systems even when Nvidia is not part of the training stack.
The open verification problem
The strongest claims around LongCat-2.0 still depend on access and confirmation. Because the model is not yet available on HuggingFace, outside researchers cannot easily test it under their own conditions. That limits confidence in how broadly the reported benchmark results translate beyond Meituan’s own presentation.
There are also hardware questions. The training cluster is described as using more than 50,000 domestically made AI ASICs, but the chip maker is unnamed. That missing detail makes it harder to evaluate how reusable, scalable, or broadly available the underlying infrastructure may be.
Even with those caveats, LongCat-2.0 gives the AI industry a concrete case to watch. It combines a 1.6 trillion parameter model, a very large domestic chip cluster, and competitive benchmark results in selected areas. The next step is whether the model becomes available for broader testing and whether its reported strengths hold up outside the original announcement.