Meituan’s LongCat-2.0 and the new reality of domestic AI training

Meituan’s LongCat-2.0 is notable not just because it is large, but because of where and how it was trained. The company says the 1.6 trillion-parameter model ran entirely on Chinese-made chips, with no Nvidia hardware involved, on a cluster of more than 50,000 domestically produced AI ASICs. It was trained on more than 35 trillion tokens. That combination makes it more than another incremental model release: it is a demonstration that frontier-scale training no longer appears to require a Western accelerator stack.

That is the strategic signal Washington will notice first. Since export controls began tightening in 2022, one of the central questions has been whether Chinese AI labs could keep scaling up large models without access to Nvidia’s best training hardware. LongCat-2.0 does not settle that debate on its own, but it does narrow the space for easy assumptions. Meituan is effectively arguing that domestic compute is now sufficient for a model of this size, at least for one training run.

The benchmark picture is more mixed than the headline might suggest, which is exactly why it matters. On SWE-bench Pro, LongCat-2.0 posts 59.5, and on SWE-bench Multilingual it scores 77.3. In those comparisons, The Decoder reports that it tops Gemini 3.1 Pro and GPT-5.5 on those specific tasks, while still falling short of Claude Opus 4.7 and 4.8. That is a meaningful result: not proof of broad superiority, but evidence that domestic hardware training can produce a model competitive enough to win on some software-engineering and multilingual workloads.

The rest of the evaluation stack tempers the enthusiasm. On IFEval, LongCat-2.0 scores 90.0; on IMO-AnswerBench, 81.8; and on GPQA-diamond, 88.9. In those tests, it trails Gemini and GPT-5.5 by wide margins in some cases. For technical readers, that pattern is familiar: a model can look unusually strong on one slice of the benchmark suite and less persuasive on another. What matters is not a single leaderboard position, but the breadth of the signal. Here the signal is real, but uneven.

That unevenness is important because independent verification remains limited. Meituan has not made the model available on HuggingFace, which means outside researchers cannot yet reproduce the reported results or inspect the system under the same conditions. The absence of public weights or an easily accessible checkpoint does not invalidate the claim, but it does mean the evidence is still company-mediated. In a field where benchmark claims are only as strong as their reproducibility, that is a material limitation.

Even with that caveat, the hardware implication is hard to ignore. A trillion-plus-parameter model trained on domestic accelerators changes the framing for Chinese chipmakers, because it suggests there is now a credible customer and use case at the top end of the market. If domestic ASICs can support LongCat-2.0-class training, then the relevant question is no longer only whether Chinese hardware can substitute for Nvidia in theory, but how quickly the software ecosystem, compiler stack, networking, and distributed-training tooling can mature around it.

That is also where Meituan’s position becomes interesting. The LongCat team has only existed since 2023, and its first model shipped late last year. In other words, this is not the product of a decade-long, deeply entrenched frontier-lab culture; it is a newer effort moving quickly through the stack. For Meituan, a company better known for consumer internet services than foundational model research, the release reads as both technical proof point and strategic messaging: the company can participate in the AI race not just as a consumer of models, but as a builder of infrastructure-adjacent capability.

The market implication is less about immediate displacement and more about bargaining power. If domestic-hardware training is viable at this scale, then Chinese AI buyers, cloud operators, and chip designers have a stronger basis for planning around local supply. That does not erase the performance gap that still appears on some benchmarks, nor does it eliminate dependence on mature tooling ecosystems that remain more developed around Nvidia. But it does suggest that the procurement logic for large-model training in China may become more locally anchored over time.

For readers looking for the next proof point, three things matter most. First, whether the model becomes publicly available for independent testing. Second, whether the reported benchmark mix holds up under broader scrutiny rather than a narrow release set. Third, whether other teams can replicate the result on similar domestic clusters, which would tell us more about the underlying ecosystem than any one model release can.

LongCat-2.0 is therefore best read as a credible milestone rather than a final verdict. It does not prove that domestic AI hardware has closed the gap with Nvidia. It does show that, under current export-controls-era conditions, a Chinese company can train a very large model on Chinese-made chips alone and produce results that are competitive enough to matter. That is a substantive shift in what is considered technically plausible, and it is exactly the kind of shift that tends to matter before it becomes obvious in market share data.