China's First Self-Developed Trillion-Parameter Large Model for Card Training and Inference Officially Open Source - 36 Kr

Meituan has open-sourced LongCat-2.0, a 1.6-trillion-parameter model trained on a 'domestic' 50,000-card cluster. Because nothing says 'trust us, we're just sharing knowledge' quite like a model that costs more to run than a small country's GDP. The model activates only 48 billion parameters per inference, which is about as thoughtful as claiming a 747 is fuel-efficient because it only uses a few engines at cruising altitude. The phrase 'self-developed' is doing a lot of heavy lifting here, implying independence from foreign chips while conveniently ignoring that the cluster is still built on GPUs that trace back to ARM architecture.

This is the sort of open-source release that makes Western regulators sweat and Eastern propagandists cheer, all while the actual carbon footprint of training this beast could power a data centre in Shenzhen for a year. The prior coverage of Chinese models quietly infiltrating U.S. markets now looks quaint — they're not just undercutting on price, they're giving away the whole farm. The 'domestic cluster' claim is particularly rich: it's the AI equivalent of 'my dog ate my homework' meets 'made in China with pride.'

And what do we get for all this compute? A model that, if the trends hold, will soon be fine-tuned for everything from cat memes to autonomous drone swarms. Meituan, a food delivery company, now holds the keys to a trillion-parameter beast. The open-source community will surely use it responsibly — just as they've done with every previous model. We'll now enjoy the benefits of Chinese state-adjacent companies pushing open-source models that are too big to run on anything but China's own hardware. Great.