On June 30, Meituan — yes, the food-delivery company — open-sourced LongCat-2.0, a 1.6-trillion-parameter model built for agentic coding. The parameter count is not the flex. The flex is that Meituan says it trained and serves the thing entirely on domestically produced Chinese chips, on a 50,000-card cluster, with zero restricted Nvidia hardware in the loop.
How it works
LongCat-2.0 is a mixture-of-experts model, so it does not fire all 1.6 trillion parameters on every token — it activates roughly 33 to 56 billion at a time, which is what makes serving it remotely affordable. It ships with a one-million-token context window and, crucially, an open-weights license. Meituan is calling it the first trillion-parameter model to complete both training and inference on domestic hardware. Even discounting the marketing, that is a milestone with teeth.
The AI tool stack actually worth paying for
One email a week. The models, tools and moves that matter, stripped of hype and filtered so you don’t have to drink from the firehose. Free, and you can bail anytime.
Why this dents the narrative
Washington’s entire AI-containment strategy since 2022 rests on one assumption: choke China’s access to top-end chips and you cap how big and capable their models can get. LongCat-2.0 is a public argument that the ceiling is higher than the export-control crowd hoped. Efficiency work, domestic accelerators, and MoE architectures are chipping away at the premium that Nvidia’s best silicon was supposed to guarantee. It does not mean the gap closed. It means the gap is a moving target, and “just deny them the chips” is aging badly.
The practical read
For anyone actually building, an open-weights trillion-parameter coding model you can self-host is a real option, not a press release — with the usual asterisks about data governance and where you are comfortable running it. The geopolitics will get the headlines. The open license is what changes your options this quarter.