A Food Delivery Company Just Open Sourced a 1.6 Trillion Parameter AI and Says It Never Touched an Nvidia Chip

Meituan, best known for delivering food across China, open sourced LongCat-2.0, a 1.6 trillion parameter model with a one million token context window, and dropped the mic on the way out. The headline claim is the spicy one: that it completed both full pre-training and inference entirely on domestic, non-Nvidia Chinese hardware, reportedly a cluster of around 50,000 domestic AI chips linked to Huawei's stack. If true, that is a very large crack in a very expensive wall.

Under the hood it is a sparse mixture-of-experts design, the same efficiency trick DeepSeek and Mixtral use, where a router picks a few expert sub-networks per token instead of firing the whole model every time. Meituan published benchmarks putting it up against Gemini, GPT-5.5 and Claude Opus, and released the weights openly, framing the whole thing as a direct answer to United States chip export controls. A delivery app doing geopolitics is peak 2026.

Now the asterisk, because Sally reads the fine print. The trained-entirely-on-domestic-chips claim and every the-benchmarks-are-insane number come from Meituan itself, not an independent lab. So treat the swagger as marketing until someone reproduces it. But even discounted, an openly released 1.6 trillion parameter, million token model is a real gift to researchers, and a real headache for anyone who bet that cutting off silicon would freeze Chinese frontier training.

• •Meituan open sourced LongCat-2.0, a 1.6 trillion parameter model with a one million token context window, on June 30, 2026.
• •It claims to be the first trillion-parameter model fully pre-trained and run on domestic, non-Nvidia Chinese hardware, reportedly a roughly 50,000 chip cluster tied to Huawei technology.
• •The model uses a sparse mixture-of-experts architecture, like DeepSeek and Mixtral, routing each token to a few expert sub-networks.
• •Meituan published benchmarks positioning it against Google's Gemini, OpenAI's GPT-5.5 and Anthropic's Claude Opus, and released the weights openly.
• •The release was framed explicitly as a response to United States export controls on advanced chips; the domestic-hardware and benchmark claims are Meituan's own and not independently verified.