LongCat 2.0 — Meituan's 1.6T model trained without NVIDIA chips

The United States has spent years and tens of billions of dollars all to stop China from building AI. But that was just proven irrelevant by a food delivery app. I'm Gigashan, I spent years living in China so it brings a smile to my face when I get to talk about Maytwan. Maytwan, why am I so much hot or quiet? That just means Maytwan delivers everything super fast and they're wildly popular. Their mascot probably helps, have a look. Who doesn't like a kangaroo? But today we're talking about cats, long cat 2.0 to be exact. It's a 1.6 trillion parameter AI model with a million token context window. These guys like animals. They originally released this model as owl alpha on open router. They let dev slam it with API calls and watch their token volume rise all the way to the global top three and then they revealed who they actually were. And they trained it entirely on domestic Chinese chips. No Nvidia at all, let me explain. When you train a model this big, you split it across thousands of chips. Every forward pass, every backward pass, the chips have to exchange massive amounts of data. Perfectly, synchronously. This is where something called NCCL comes in. That stands for Nvidia Collective Communications Library, but wait, I thought there was no Nvidia. We'll get to that. For now just know it's the nervous system of the GPU cluster. A chip sends its gradient to other chips. This is the mathematical adjustment formula or the direction and size of the error correction. Here's…