Meituan’s LongCat 2.0 runs on chips the US can’t export and it just beat GPT-5.5

LongCat 2.0 has 1.6 trillion parameters, processes a million-token context window, and just outperformed OpenAI‘s GPT-5.5 on the leading benchmark for real-world software engineering. It was built by Meituan — a company most people outside China know, if at all, as a food delivery app.

The model scored 59.5 on SWE-bench Pro against GPT-5.5’s 58.6. That margin might look small, but the context of how this system was built is what makes the achievement striking. Meituan trained LongCat 2.0 on more than 50,000 domestic Chinese AI accelerator chips — not Nvidia hardware, which China has been unable to import in quantity since 2022. The manufacturer hasn’t been confirmed, but the technical community’s best guess is Huawei’s Ascend 910B series. If accurate, it would represent the first time a Huawei-trained model has reached frontier performance in transparent public benchmarks.

LongCat 2.0 uses a Mixture-of-Experts architecture, a design that activates only a fraction of the total network for any given task. Between 33 and 56 billion parameters fire per query, which keeps the system fast and efficient without requiring hardware that most developers have access to. The one-million-token context window is what makes it particularly interesting for coding work: the model can hold an entire large codebase in memory at once, reading across files and functions without losing context. On SWE-bench Multilingual, it scored 77.3. On Terminal-Bench — which tests an AI’s ability to navigate an actual Unix environment — it reached 70.8.

Before the official announcement, the model ran quietly on OpenRouter under the name “Owl Alpha” for roughly two months. Developers who used it noticed its unusual coding ability but didn’t know where it came from. Meituan confirmed after release that Owl Alpha was LongCat 2.0 in an earlier evaluation stage.

The MIT license tells only part of the story. Despite being one of the most open and commercially permissive licenses available, the model weights haven’t been released yet. The GitHub repository and the Hugging Face model card both say they’re “coming soon.” Right now, developers can access LongCat 2.0 through a hosted API, which means querying it remotely rather than running it locally. Self-hosting, fine-tuning, and downloading the weights aren’t possible yet.

What Meituan has built is worth the anticipation. The company processes over 50 million food and grocery orders daily and quietly built an internal language model called Zhichi before turning to external AI development. LongCat 2.0 is its first public model and, by benchmark performance, one of the most capable coding systems available under an open license.

When the weights arrive, LongCat 2.0 enters direct comparison with Meta’s Llama 4.1 and DeepSeek-V4 Pro as candidates for the most capable open coding model a developer can download and run. On the hardware question, the confirmation of which domestic chips produced these results — whenever it comes — will matter well beyond the AI community.

Tags: artificial intelligence, Huawei, Open Source, GPT-5.5, LongCat 2.0, Meituan