Meituan Open-Sources LongCat-2.0, Trained on Chinese Chips

Meituan, the Chinese food-delivery and services giant better known for moving takeout than training frontier AI, released and open-sourced its next-generation LongCat-2.0 model on Tuesday, and made a claim that landed harder than the model's size. The company says the 1.6-trillion-parameter system is the first of its scale trained and served entirely on domestically produced Chinese chips, from the brutal work of pre-training through everyday inference.

The specifications alone would draw attention. LongCat-2.0 carries 1.6 trillion parameters, uses a Mixture-of-Experts design that activates only a slice of the network for any given task, and ships with a one-million-token context window. Meituan released the weights under a permissive MIT license and posted them to Hugging Face, putting the model in the hands of anyone who wants to run it or pull it apart. In Meituan's own framing, it is "the industry's first trillion-parameter model to complete end-to-end training and inference on a 50,000-chip domestic compute cluster."

Why the hardware claim matters more than the parameter count

Plenty of Chinese labs already run inference, the comparatively light task of answering a query, on home-grown silicon. Pre-training is the expensive part. It is the phase where a model chews through enormous data sets to learn its basic patterns, and it is exactly where the most advanced accelerators have mattered most. That is why Meituan's insistence that LongCat-2.0 was both pre-trained and served without Nvidia is the headline, not the trillion-plus parameters.

The training reportedly ran on a cluster of roughly 50,000 domestic accelerators, coordinated through Huawei's Collective Communication Library, the software layer that keeps thousands of chips working in sync. No Nvidia H100s. No AMD accelerators. If the account holds up, it speaks directly to the strategic question hanging over China's AI sector: whether the country can build models at frontier scale without American hardware at all.

That question exists because Washington has spent the last several years restricting exports of the most capable chips on national-security grounds. Beijing has responded by pouring money into a domestic stack, from chip design to fabrication to the cluster software that ties it together. LongCat-2.0 is the software proof-of-concept for that push, a large, public model meant to show the domestic hardware can carry a training run that used to require restricted silicon.

The performance claims, and the caveats

Meituan is not pitching this as a curiosity. The company says LongCat-2.0's performance is comparable to Google's Gemini 3.1 Pro, and that it scored 59.5 on the SWE-bench Pro coding benchmark, edging past a reported 58.6 for GPT-5.5. VentureBeat described the release as a "near-frontier agentic coding model," and reported that the system is the engine behind Owl Alpha, an anonymous model that had been quietly topping developer usage charts on OpenRouter before its origin was disclosed.

Those numbers deserve a skeptic's eye. They come from Meituan, and independent evaluators have not yet confirmed them. The hardware claim is even harder for outsiders to verify directly, since it rests on the company's account of its own infrastructure. What makes this different from a pure marketing exercise is that the weights are public. The open-source community can now run LongCat-2.0 against the same benchmarks and decide for itself whether the model matches the marketing. That verification loop, playing out over the coming weeks, is the part worth watching.

What it means for founders and operators

For anyone building a company on top of AI, the interesting layer here is not geopolitics, it is the supply of capable open models. A 1.6-trillion-parameter model tuned for agentic coding, released under a license that permits commercial use, lowers the floor on what a startup can self-host or fine-tune without paying frontier-lab API rates. If the coding benchmarks survive scrutiny, LongCat-2.0 becomes another serious option for teams that want to run inference on their own terms, control their data, and avoid vendor lock-in.

There is a second signal for operators who track the compute supply chain. Every model trained without Nvidia narrows the practical gap that export controls were designed to widen, and it hints at a future where the hardware underneath your AI stack is more contested and more varied than the Nvidia-by-default present. That matters for pricing, for availability during shortages, and for the risk calculus of building on any single vendor.

The caution flag is trust. Model provenance, security review of open weights, and the durability of a benchmark lead all become due-diligence items when the model comes from a large consumer company rather than a dedicated lab. Founders adopting LongCat-2.0 will want to treat the performance claims as a starting hypothesis to test against their own workloads, not a settled fact.

Meituan is an unlikely flag-bearer, and that is part of the story too. It runs one of the world's largest on-demand logistics operations, where routing, demand forecasting, and customer service all run on compute. A model trained on domestically secured silicon insulates that compute from the next turn of the export-control screw, which makes the appeal concrete rather than symbolic. Open-sourcing it seeds adoption among developers and signals confidence that the underlying chips can keep pace.

Frequently asked questions

What is LongCat-2.0?

LongCat-2.0 is an open-source large language model released by Meituan on June 30, 2026. It has 1.6 trillion parameters, a Mixture-of-Experts architecture, and a one-million-token context window, and is tuned for agentic coding and reasoning tasks. The weights are available on Hugging Face under an MIT license.

Why is it significant that it was trained on Chinese chips?

Meituan says the model was both pre-trained and run on a cluster of roughly 50,000 domestic accelerators, without Nvidia hardware. Pre-training is the most compute-intensive stage of building a model, so doing it entirely on home-grown silicon suggests China can reach frontier scale despite US export controls on advanced chips.

How good is it compared with US models?

Meituan claims performance comparable to Google's Gemini 3.1 Pro and a SWE-bench Pro coding score that edges past GPT-5.5. Those figures come from Meituan and have not been independently verified. Because the weights are public, outside researchers can now test the claims directly.

Can startups actually use it?

Yes. The MIT license permits commercial use, and the weights are downloadable, so teams can self-host or fine-tune the model rather than relying on a paid API. As with any open model, adopters should review security, provenance, and real-world performance against their own workloads before deploying.

What should operators watch next?

The key signals are independent benchmark results from the open-source community, any confirmation or challenge to the domestic-hardware claim, and how quickly developers adopt the model in production. Together those will show whether LongCat-2.0 is a durable option or a headline that fades.