For years, the working assumption in tech policy circles has been blunt: no NVIDIA chips, no frontier AI. The export con

For years, the working assumption in tech policy circles has been blunt: no NVIDIA chips, no frontier AI. The export controls on high-end GPUs like the H100 and the newer Blackwell line were designed precisely around that bottleneck. Cut off access to the best silicon, the theory went, and you slow a country’s ability to train the largest, most capable models. Meituan just called that bluff.

The Chinese tech giant, best known for dominating food delivery and local services, has released LongCat 2.0. It is a 1.6-trillion-parameter model built on a Mixture-of-Experts architecture. The kicker? The team trained it entirely on domestic Chinese chips. No NVIDIA H100s. No Blackwell GPUs. And rather than hiding it behind an API, Meituan has dropped the weights, training code, and full data pipeline on GitHub for anyone to inspect.

What LongCat 2.0 Actually Brings

Let’s look at the hardware-agnostic facts first. At 1.6 trillion parameters, LongCat 2.0 sits among the largest open-weight models ever released. Parameters are not the only measure of quality, but at this scale, they signal serious engineering ambition. Because it uses a Mixture-of-Experts design, only a subset of those parameters activates for any given task. That keeps inference costs from ballooning out of control while still allowing the model to store an enormous range of knowledge and reasoning patterns.

The context window hits one million tokens. That matches the high-end offerings from OpenAI and Anthropic, and it changes what the model can practically do. With a window that large, you can feed it entire legal contracts, months of chat logs, or vast code repositories in a single prompt. For developers building applications that require memory across long documents, this is not a marginal improvement. It is a functional necessity.

Then there is the openness. Meituan did not release a sanitized demo or a restricted API. The model weights are available, the training code is public, and so is the pipeline. That matters for researchers who want to reproduce results, for enterprises that need to audit behavior, and for engineers who need to fine-tune the model on proprietary data without sending anything to a third-party server.

The Hardware Story Everyone Missed

The headline here is not just the parameter count. It is the silicon underneath. LongCat 2.0 was trained on domestic accelerators, specifically chips like the Huawei Ascend series. That is a radically different proposition from slapping an existing framework onto an NVIDIA cluster and hitting run.

Training a trillion-parameter model requires solving distributed computing problems at the extreme edge. Memory bandwidth, inter-chip communication, and floating-point performance all have to be managed with ferocious precision. NVIDIA’s advantage has never been just the raw GPUs. It is the CUDA ecosystem, the optimized kernels, and the collective knowledge of how to squeeze performance out of that stack. Building a comparable model on Ascend hardware means Meituan’s engineers had to do the grueling work of adapting training frameworks, rewriting low-level operations, and debugging distributed training runs on a fundamentally different architecture.

That they succeeded suggests something larger than one model. It shows that Chinese hardware-software co-design is maturing to the point where the absence of Western chips is no longer a hard stop. It is a constraint, and an expensive one, but not an impossible barrier.

Where It Actually Performs

LongCat 2.0 scores strongly in Chinese reasoning, mathematics, and long-context retrieval. Those are specific, high-value benchmarks. Mathematical reasoning tests a model’s ability to handle logic and symbolic manipulation. Long-context retrieval tests whether the model can find a needle of information in a million-token haystack without losing track. Passing both is the difference between a model that sounds smart and one that can actually do work.

Meituan is positioning this as a direct alternative to Llama, DeepSeek, and Qwen. For Chinese language tasks, that competition is especially sharp. Models trained predominantly on English internet data often stumble on classical Chinese references, local regulatory language, financial terminology used in mainland markets, and the informal shorthand that dominates Chinese social media. A model built with deep Chinese language skills, fed by a company whose core business runs on Chinese consumer behavior, has a structural advantage in those scenarios.

That matters for more than chatbots. Legaltech firms analyzing PRC contract law, researchers parsing historical texts, banks processing Mandarin loan documentation, and customer service platforms handling regional dialects all need models that understand nuance rather than translating through an English-centric lens.

Meituan’s Hidden Weapon: Scale and Data

Meituan is not a research lab with a delivery app on the side. It is an operational juggernaut that coordinates millions of riders, restaurants, and merchants across China every day. That scale generates a torrent of real-world language data. Customer complaints routed to chatbots. Restaurant descriptions that blend regional slang with formal business registration details. Routing instructions that mix address systems with informal landmarks. Merchant service tickets that touch on taxes, hygiene regulations, and local ordinances.

This data is messy, contextual, and deeply local in a way that generic web crawl data can never replicate. Feeding that into LongCat 2.0 gives the model a practical grounding that purely academic training runs lack. It is one thing to train on polished Wikipedia text. It is another to train on the chaotic, transactional language of actual commerce.

Why the Ground Is Shifting

If companies can train frontier-class models on domestic silicon, the entire strategic logic behind chip export bans begins to erode. The bans were built on the premise that controlling NVIDIA supply lines would control AI capability. That premise assumed no viable alternative ecosystem existed.

LongCat 2.0 is not a declaration that Chinese chips have matched NVIDIA on every metric. It is proof that they do not need to match perfectly to get the job done. Sufficient memory, sufficient bandwidth, and sufficiently smart software optimization can close enough of the gap to produce competitive results. That is a much lower bar than total parity, and it is a bar that appears to have been cleared.

For global AI supply chains, the implication is clear. The assumption that all serious training runs must flow through NVIDIA hardware is now dead. That shifts power toward national sovereignty in AI development. Countries and companies watching from the sidelines no longer see a single chokepoint supplier as the only path to frontier capability. They see a bifurcation, and possibly a multipolar hardware landscape, emerging faster than most predicted.

The Real Takeaway

LongCat 2.0 is more than a technical release. It is a test of a political hypothesis, and the hypothesis just failed. Meituan has demonstrated that a consumer tech company with the right data, the right engineering team, and the right hardware can train a 1.6-trillion-parameter open model without touching a single banned Western GPU.

For developers, this means a new open-weight option with genuine depth in Chinese language and long-context tasks. For policymakers, it means sanctions based on hardware denial must account for adaptation, not just acquisition. For the rest of the industry, it means the map of who can build what, and with which tools, is being redrawn in real time.

The export bans may have bought time. They also appear to have bought an alternative.