Meituan is the company that brings you dinner in China. It’s also, apparently, now in the business of releasing some of the largest open-source AI models on the planet. On June 30, Meituan open-sourced LongCat-2.0, and there’s a genuinely wild backstory behind how this model showed up.

Wait, It Was Already Out There?

Here’s the part that makes this release different from a normal launch: LongCat-2.0 wasn’t a total surprise. For two months before Meituan revealed it, this exact model had been quietly running on OpenRouter under a codename, “Owl Alpha,” and climbing to the top of global developer usage charts without anyone knowing who made it.

So people had already been using this model, ranking it, and comparing it to others, all without knowing it came from a food delivery company. Meituan just pulled back the curtain and said, “yeah, that was us.”

The Specs

LongCat-2.0 is a Mixture-of-Experts (MoE) model with 1.6 trillion total parameters. If that number sounds enormous, here’s the catch that makes it manageable: it doesn’t use all 1.6 trillion parameters for every single request. Instead, it dynamically activates somewhere between 33 and 56 billion parameters per token, depending on how complex the task is.

Think of it like a company with 1.6 trillion employees on payroll, but for any given task, only a small specialized team of a few dozen billion “employees” gets called in to work on it. That’s what keeps a model this large from being impossibly expensive to run.

Other specs worth mentioning:

• A native 1 million token context window, meaning it can take in something the size of an entire large codebase and reason about all of it at once
• Released under the MIT license, which is about as permissive as open-source licenses get. Companies can take LongCat-2.0, modify it, and bake it directly into commercial products without the legal restrictions that come with stricter open-source licenses
• Built specifically for agentic coding: understanding, writing, and executing code as part of automated workflows, not just answering questions about code

No Nvidia Involved, At All

This might be the most notable part of the whole story. Meituan trained LongCat-2.0 on a 50,000-chip cluster made entirely of domestic Chinese hardware, widely believed to be Huawei’s Ascend 910 chips. No Nvidia GPUs, no AMD chips, nothing from outside China touched this training run.

That matters because of the ongoing US export restrictions on advanced AI chips to China, which have been in place since late 2022 and have only gotten stricter since. The fact that a trillion-parameter model came out of a fully domestic hardware pipeline is being treated as a real signal that those restrictions haven’t stopped Chinese labs from reaching frontier-scale training, just changed the tools they’re using to get there.

How Good Is It, Really?

On benchmarks Meituan reported itself, LongCat-2.0 scored 59.5 on SWE-bench Pro, which is a test of how well a model handles realistic software engineering tasks. For comparison, that edges out GPT-5.5’s 58.6 and Gemini 3.1 Pro’s 54.2 on the same test. It also posted a 70.8 on Terminal-Bench 2.1, a benchmark for command-line and tool-use tasks.

Meituan describes the model’s overall performance as comparable to Gemini 3.1 Pro. That said, these numbers come from Meituan’s own testing, so they haven’t gone through independent verification yet the way established benchmarks from neutral third parties would provide. It’s also worth noting other reports suggest LongCat-2.0 still lags behind Claude’s Opus models on broader agent tasks, even with the strong SWE-bench Pro number.

Where the Open-Source Part Gets a Little Fuzzy

Even though this is billed as an open-source release, not everything is fully public yet. The model is available to try on longcat.ai and OpenRouter, and it’s currently one of the top three most-used models on OpenRouter by call volume. But the full model weights are still listed as pending in some of the coverage around the release, so “open-source” here is more of a spectrum than a single clean checkbox at the moment.

If you want to try it yourself without waiting on anything, the easiest path is through OpenRouter, since it’s already been running there under its old alias for months with real usage behind it.

LongCat-2.0 is a solid example of how much the AI landscape has widened this year. It’s not just OpenAI, Anthropic, and Google racing at the top anymore, it’s a food delivery app in China quietly running one of the most-used coding models in the world for two months before anyone knew its name.