Meituan's LongCat-2.0: the first trillion-parameter model built without Nvidia

The number Meituan wants you to read is 1.6 trillion parameters. The number that actually reshapes the competitive map is 50,000 — the count of domestic ASICs in the cluster that trained and now serves LongCat-2.0, with, per Meituan, not a single Nvidia GPU in the loop. If that claim holds up to independent scrutiny, LongCat-2.0 is the first trillion-parameter model built end-to-end on Chinese silicon, and that is a supply-chain fact with far longer legs than any benchmark row.

In late June 2026 Meituan — China's largest food-delivery and local-services platform, not a name most Western developers associate with frontier models — open-sourced LongCat-2.0 under a permissive MIT licence. The core specifications:

The dynamic activation band — the model routes each token through somewhere between roughly 33 and 56 billion parameters rather than a fixed slice — is the efficiency lever that makes a 1.6T-parameter model economical to run. One caveat at launch: Meituan listed the full weights as 'coming soon' rather than all posted, so the open-weight claim was, at announcement, partly a promise.

For two years the working assumption in Western AI policy has been that US export controls on advanced Nvidia accelerators impose a hard ceiling on how large a model China can train. LongCat-2.0 is the first concrete counter-example at the trillion-parameter scale. Meituan's claim is not that domestic ASICs match an H100 chip-for-chip — it is that a cluster of more than 50,000 of them, engineered around, was sufficient to train and serve a near-frontier model end-to-end. That reframes the export-control question from 'can they train a big model?' to 'how much does the domestic-silicon detour actually cost them in time and efficiency?' — a narrower and more uncomfortable question.

Meituan positions LongCat-2.0 as a near-frontier agentic coding model, and the figures it published are strong — but every one of them is vendor self-reported and not yet independently verified:

The widely repeated 'beats GPT-5.5' line is technically true and materially thin: it rests on a single benchmark, SWE-Bench Pro, by a margin of under one point (59.5 vs 58.6) — inside the noise band of most eval harnesses. Treat 59.5 as a claimed ceiling under favourable conditions, not a settled ranking. Independent scorers such as SWE-bench maintainers, Artificial Analysis or LMArena had not published verified numbers at announcement.

LongCat-2.0 does not arrive in isolation. It is the latest and heaviest entry in a run of Chinese open-weight releases — DeepSeek, Moonshot, MiniMax and now Meituan — that have quietly captured the demand side of the market: Chinese open-weight models now account for roughly 61% of OpenRouter's top-10 traffic. The 'Owl Alpha' episode is the tell. A model can lead a developer usage chart for two months entirely on merit and price before anyone knows who built it, because open weights and an OpenRouter endpoint remove the brand from the buying decision. That is the mechanism by which the open-weight surge converts into installed base — and installed base, not a leaderboard crown, is what compounds. The variables worth tracking now: whether the full weights ship as promised, whether independent benchmarks confirm the coding scores, and whether anyone outside Meituan can reproduce the non-Nvidia training claim.