Meituan trained a 1.6 trillion parameter model without a single Nvidia chip

A food delivery company just broke the central assumption of US AI export controls.

Meituan released LongCat-2.0 today: a 1.6 trillion parameter Mixture-of-Experts model, open-sourced on Hugging Face, trained end-to-end on a 50,000-chip cluster of domestic Chinese silicon. No Nvidia. No AMD. No workaround. Just homegrown hardware running the full pipeline from pre-training through inference.

That distinction matters more than anything else in this announcement. DeepSeek’s V4-Pro from April, the previous Chinese frontier benchmark, used domestic chips only for the lighter inference step. Pre-training is where you stream trillions of tokens through tens of thousands of accelerators and pray the cluster interconnect doesn’t crash. That’s the computationally brutal part. That’s the part everyone said China couldn’t do without Nvidia.

Meituan did it. To keep the training run stable without Nvidia’s NCCL, they used Huawei’s Collective Communication Library (HCCL) for chip-to-chip communication. The entire export control thesis rested on training being the wall China couldn’t cross. That wall just came down.

The model itself is a Mixture-of-Experts design that activates 33 to 56 billion parameters per token despite the 1.6 trillion total. That’s an efficiency adaptation to the memory and bandwidth constraints of domestic chips, which still lag Nvidia on raw specs. The context window hits 1 million tokens. On benchmarks, Meituan reports 59.5 on SWE-bench Pro and 70.8 on Terminal-Bench, both software engineering evaluation suites. Take those numbers as vendor claims until an independent leaderboard run confirms them. “On par with DeepSeek V4-Pro” headlines have a history of softening under scrutiny.

Here’s the detail that should scare Washington: LongCat-2.0 was already ranking at the top of OpenRouter’s usage charts before Meituan open-sourced it. Developers outside China were already choosing it. The model earned its reputation through real usage, not a press release.

Meituan is not an AI-native company. They deliver noodles and groceries. If a company whose core business is logistics can complete a trillion-parameter training run on domestic silicon, the question stops being whether China can train frontier models. It starts being how quickly Huawei’s training stack becomes the default for everyone inside the firewall.

DeepSeek and Zhipu, the two Chinese labs most people watch, have so far used domestic chips only for inference. Meituan just moved the line past both of them. A delivery app company. That’s either a sign of how deep China’s compute independence runs, or a signal that the bigger labs are holding back.

The reporting does not address the harder economic question: how wall-clock time, energy cost, and chip count compared to an equivalent Nvidia run. SCMP cites tens of thousands of ASICs but puts no number on cost or duration. Nobody has named which specific domestic chip line did the work. That detail would tell you whether this generalizes to every Chinese lab or just Meituan.

LongCat-2.0 is available now on Hugging Face under the meituan-longcat organization. The lineage goes LongCat-Flash (560 billion parameters, September 2025) to LongCat-Next (multimodal, March 2026) to LongCat-2.0 (today). Nearly triple the parameter count in nine months, all on domestic hardware.