What is LongCat-2.0? A giant free AI from China, built without Nvidia

If you saw headlines about a Chinese food-delivery company releasing a giant free AI model and did a double-take — you read that right. In late June 2026, Meituan (think of it as China's Uber Eats plus much more) open-sourced an AI called LongCat-2.0. And it's not a small experiment: it's one of the largest AI models ever released for free. Here's what it is and why people are paying attention, without the jargon.

LongCat-2.0 has 1.6 trillion parameters — parameters are roughly the learned patterns that make a model smart, and 1.6 trillion is enormous. But here's the clever bit: it uses a design called 'Mixture-of-Experts', which means it only switches on the part it needs for each word (somewhere between about 33 and 56 billion parameters at a time) instead of the whole thing. Picture a huge library where you only ever open the one shelf that has your answer — you get the depth of the big library without the cost of reading every book. That's how a model this large stays affordable to run.

It also has a 1-million-token context window. In everyday terms, that's roughly 750,000 words — about ten full-length novels — that it can hold in mind at once. So you could hand it an entire codebase, a year of emails, or several long reports in a single go and ask questions across all of it. That kind of memory is where models like this genuinely shine.

Here's the detail that makes this more than just another big model. Almost every powerful AI you've heard of was trained on chips made by Nvidia, a US company. Because of US export rules, China can't freely buy Nvidia's most advanced chips. Meituan says it trained and ran LongCat-2.0 entirely on Chinese-made chips instead — more than 50,000 of them — with no Nvidia hardware at all. If that's accurate, it's the first model this size built without Nvidia, which matters because it suggests those export rules may not hold China back as much as was assumed. It's a bit like a country that was cut off from imported car parts building a working car entirely from its own factories.

Meituan says LongCat-2.0 is excellent at coding tasks, and even points to one test (called SWE-Bench Pro) where it edges out OpenAI's GPT-5.5, scoring 59.5 to 58.6. Sounds great — but two honest caveats. First, that lead is tiny (less than a point) and it's only one test out of many. Second, and more important: those scores come from Meituan itself and haven't been checked by independent testers yet. So it's fair to say LongCat-2.0 looks genuinely strong, but not yet fair to say it's proven to beat GPT-5.5. One more practical note: when it was announced, not all of the model files were posted yet — Meituan listed the full weights as 'coming soon' — so check what's actually available before you plan to use it.

Should you care as an everyday user or builder? If you write code or build AI products, yes — a capable model you can run yourself, for free, changes the economics, and this is part of a wave: Chinese open models now make up roughly 61% of the traffic on OpenRouter's top-10 list. The fact that developers were already flocking to it as the mystery 'Owl Alpha' before anyone knew Meituan made it tells you it holds up in real use. If you just chat with AI casually, this won't change your day directly — but it's a sign of how fast, and how globally, capable AI is spreading.