News · 2026-07-11

Meituan open-sources LongCat-2.0, a trillion-parameter model it says was trained end-to-end on Chinese chips

Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter mixture-of-experts model, and made a claim that landed harder than the benchmarks: it says the model was both trained and deployed end-to-end on domestic Chinese AI chips, with no Nvidia hardware anywhere in the pipeline. The weights are released under the permissive MIT license, and for two months before the reveal the same model had been running anonymously on OpenRouter as 'Owl Alpha,' quietly climbing to the top of developer usage charts.

Key facts

• Scale: 1.6 trillion total parameters, about 48 billion active per token (a mixture-of-experts design), plus a separate 135-billion-parameter N-gram embedding.
• Hardware claim: trained and served on 'AI ASIC superpods'; secondary reports name roughly 50,000 chips from Huawei, Moore Threads, and MetaX.
• When and who: released mid-July 2026 by Meituan (yes, the food-delivery company), under an MIT license.
• Primary source: the LongCat-2.0 GitHub repository.

Start with what the model is, because the food-delivery detail keeps stealing the headline. LongCat-2.0 is a large open-weight model tuned for coding and agentic work -- the kind of multi-step, tool-using tasks where a model reads a codebase, plans, and edits. It uses a mixture-of-experts layout, meaning only a small slice of its enormous parameter count fires for any given token, so it stays affordable to run despite the trillion-plus headline number. On top of that, Meituan bolts on a 135-billion-parameter 'N-gram embedding' -- essentially a cache of common statistical patterns (function signatures, import boilerplate, repeated code structures) so the expert layers can spend their capacity on the novel logic of a prompt rather than re-learning patterns they have seen thousands of times.

The headline-grabbing part is the hardware. Meituan's own text uses the careful phrase 'AI ASIC superpods' and 'alternative hardware platform,' and states the run had 'no rollbacks or irrecoverable loss spikes' across millions of accelerator-days -- a pointed jab at the instability usually blamed on non-Nvidia software stacks. Secondary reporting fills in what Meituan won't say directly. Decrypt frames the significance sharply: 'This is the first trillion-parameter model trained and deployed end-to-end on domestic Chinese ASICs, not just served on them after training elsewhere.' By contrast, DeepSeek's earlier models used Huawei chips only for inference while pretraining ran on Nvidia hardware. If LongCat's claim holds, it is a milestone for China's push to build an AI stack that doesn't depend on export-controlled American GPUs.

The caveat matters, and it is the honest center of this story. No independent party has verified the hardware composition. As one hands-on reviewer put it, 'the claim is sourced from Meituan's official announcement... no independent verification of the hardware composition has been published.' The performance claims deserve the same skepticism: LongCat-2.0's benchmarks are self-reported, and the widely circulated 'beats GPT-5.5' framing overstates them. The real picture is that it edges GPT-5.5 on one agentic-coding test and a math-answer test, and trails on terminal tasks, browsing, instruction-following, and graduate-level science. It is genuinely competitive, and against the newest closed frontier models it loses more than it wins.

Why it matters anyway comes down to price and openness. LongCat-2.0 costs a fraction of the closed frontier -- roughly $0.75 per million input tokens against GPT-5.5's $5 -- and the weights are MIT-licensed, so anyone can self-host and sidestep the data-jurisdiction concerns that come with routing prompts through Meituan's servers. Developers who tried it during the Owl Alpha period consistently mistook it for a next-generation GLM or a Kimi variant; nobody guessed a food-delivery giant. That anonymous-launch-then-reveal move -- proving the model in the wild before attaching a brand to it -- is becoming a recognizable Chinese AI playbook, and LongCat ran it longer and louder than anyone before. The technology is real; the marketing is shrewd; and the biggest claim is still, for now, taken on Meituan's word.

Primary source, verified: read the paper →

Ground Truth. (2026, July 11). Meituan open-sources LongCat-2.0, a trillion-parameter model it says was trained end-to-end on Chinese chips. Ground Truth. https://groundtruth.day/news/longcat-2-trillion-param-open-model-domestic-chips.html

@misc{groundtruth:longcat-2-trillion-param-open-model-domestic-chips,
  title  = {Meituan open-sources LongCat-2.0, a trillion-parameter model it says was trained end-to-end on Chinese chips},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/longcat-2-trillion-param-open-model-domestic-chips.html}
}

Topics: open-weights · china · moe · coding-models · hardware