A trillion-parameter model just shipped from a food delivery company
Meituan is not a dedicated AI lab. It runs one of the largest on-demand delivery and logistics operations in the world, and its LongCat research team was only founded in 2023. On June 30, the company released LongCat-2.0, a Mixture-of-Experts model with 1.6 trillion total parameters (roughly 48 billion active per token) and a 1 million token context window, and published it under an MIT license on GitHub and Hugging Face, according to Reuters, VentureBeat, and SiliconANGLE. Meituan says the model was trained and run entirely on a cluster of 50,000 domestic Chinese AI chips, without naming the chipmaker. Reuters places that inside a broader pattern: DeepSeek, Alibaba, ByteDance, and other Chinese AI developers have been working to reduce their dependence on Nvidia hardware since Washington tightened export controls on advanced chips in 2022, with domestic chipmakers Huawei and Enflame picking up share. A preview version of the model, per Meituan, had already become one of the three most-used models on the OpenRouter marketplace before the full release.
Why a claim like this travels fast, and why that is not the same as proof
Meituan also claimed LongCat-2.0 matches or exceeds several proprietary models, including Google's Gemini, OpenAI's GPT-5.5, and Anthropic's Claude Opus, on some coding and agent benchmarks. Reuters is explicit that this is Meituan's own claim. The Next Web notes that independent verification, meaning outside developers actually running the released model against those benchmarks themselves, was still pending at the time of the announcement. That distinction matters more than the headline number. A company publishing its own benchmark results and a neutral third party confirming them are two different levels of evidence. Every AI lab, American or Chinese, has an incentive to describe its own release in the best possible light at launch. The responsible way to read a story like this is: a real, verified model shipped, with real, unverified performance claims attached to it. Both parts are true at once, and a business decision should only rest on the first part until the second part is checked.
What a serious business should not misunderstand here
Three details in the reporting are easy to skip past, and each one changes what this news actually means for a business. First, size. At 1.6 trillion parameters, this is not a model most businesses could run on their own hardware even if they wanted to. SiliconANGLE notes it is built for data center or cloud-scale deployment, not for on-premises use by a typical company. "Open weight" does not mean "easy to self-host." Second, availability. RuntimeWire, reporting the same day as the launch, noted that the public GitHub repository and Hugging Face model card listed the full downloadable weights as "coming soon" at launch, even though the release itself was real and the repository was public under an MIT license. An open-source announcement and a fully usable open-source release are not always the same day. Third, what "trained on domestic chips" actually proves. It is a real and specific technical claim, since Meituan did not name the chip supplier, and it has not been independently audited. It says something about China's progress reducing Nvidia dependence. It does not, on its own, tell a business anything about whether the model is a good fit for a specific workflow, or how the data of anyone using Meituan's hosted version would be handled.
The operational lesson: evaluate the model you can test, not the model in the press release
Cheap, open, and benchmark-topping AI models are going to keep showing up, from Chinese labs and American ones. The lesson from LongCat-2.0 is not "avoid Chinese models" or "wait for the next one." It is that the evaluation checklist should stay the same regardless of who ships the model or how impressive the announcement reads. Before any business routes a real workflow through a new model, four questions matter more than the vendor's own benchmark chart. What happens on your own test cases, not the vendor's published ones. Where does the data go if you use a hosted API versus running the model yourself, and what does that vendor's data handling policy actually say. What does "open source" include in practice, full weights, training data, documentation, versus what is marketing language on release day. And does your workflow need a model this large at all, since bigger and cheaper is not automatically the right fit for a task that a much smaller model could handle at lower cost and lower complexity.
What a serious business should do next
Do not add LongCat-2.0, or any newly announced open-weight model, to a real business workflow based on the launch announcement alone. If a use case looks like a fit, the sequence is: define what the task actually needs, request the vendor's own documentation on data handling and licensing terms, run it against your own sample data and a smaller number of tests before any commitment, and only then compare cost and performance against what you use today. This applies whether the model comes from Meituan, DeepSeek, OpenAI, Anthropic, or anyone else. The nationality of the lab is not the risk factor that matters most. The gap between a benchmark claim and a workflow that has actually been tested on your business is the risk factor that matters.
The Atlacis view
Open-weight models are giving business owners more real options than they had two years ago, and that is a good thing. It also means more announcements will arrive with impressive numbers attached before anyone outside the company that built the model has checked them. Our job is not to tell you whether LongCat-2.0, or the next model like it, is good. It is to help you build a simple, repeatable way to test any new model against your own workflow and your own data before you trust a headline benchmark with something that matters to your business.
The short version
- Meituan, a Chinese food delivery and logistics company, released LongCat-2.0 on June 30, 2026, a 1.6 trillion parameter open-weight model published under an MIT license.
- Meituan says the model was trained and run entirely on a cluster of 50,000 domestic Chinese AI chips, without naming the chipmaker, as part of a broader Chinese industry push to reduce dependence on Nvidia hardware since 2022 export controls.
- Meituan also claims the model matches or beats several proprietary models, including Google Gemini, OpenAI GPT-5.5, and Anthropic Claude Opus, on some coding benchmarks. Reuters and The Next Web both note this is Meituan's own claim, not yet independently verified.
- At 1.6 trillion parameters, the model is built for data center or cloud deployment, not for typical on-premises business use, and some reporting noted the full downloadable weights were listed as 'coming soon' at launch despite the open-source announcement.
- The operational lesson is not about picking a side between American and Chinese AI labs. It is to evaluate any new model, from any vendor, against your own test cases and data handling requirements before a benchmark claim becomes a business decision.
Where ATLACIS can help
Sources
- Reuters: China's Meituan says new AI model trained on domestic chips (June 30, 2026)
- VentureBeat: Meituan open sources LongCat-2.0, the 1.6T, near-frontier agentic coding model, trained entirely on Chinese chips (June 30, 2026)
- SiliconANGLE: China's Meituan open-sources massive LongCat-2.0 AI model, saying it was trained on domestic chips (June 30, 2026)
- The Next Web: China's Meituan says its new AI model was trained on domestic chips (June 30, 2026)