Meituan Trained a Frontier Model Without Nvidia

On June 30, Meituan — the company whose main product is getting you dinner — open-sourced LongCat-2.0: 1.6 trillion parameters, a one-million-token context window, trained end-to-end on 50,000 domestic Chinese chips, no NVIDIA GPU in the stack. It is the first confirmed frontier-scale language model to complete full pre-training on domestic Chinese silicon. That milestone had been anticipated for years and denied for the same length of time. Now it exists, and it came from a food delivery company.

The significance isn’t the model’s benchmark scores. It’s what the training run proves about the structure of the export control regime — and what the export control regime can no longer honestly claim.

The Last Firewall Was Training

Pre-training is the expensive part of building a large language model. You need to feed 35 trillion tokens through a 1.6 trillion parameter model, keep gradients stable across the full run, and keep 50,000 chips communicating reliably without dropping state. The software that makes this tractable — NVIDIA’s NCCL communication library, the CUDA ecosystem, years of production engineering baked into the stack — is where US leverage lived after the chip restrictions went into effect. You could smuggle H100s. Chinese companies did: CSIS documented Huawei obtaining 2 million chiplets from TSMC through shell companies, and hundreds of millions in banned GPU-containing servers moved through informal channels. But the training stack itself — the accumulated knowledge of how to orchestrate a 50,000-card cluster stably across days-long runs — was supposed to be the chokepoint that chip possession alone couldn’t buy.

DeepSeek’s V4-Pro, released in April 2026, cracked that assumption halfway. DeepSeek ran inference on domestic chips while keeping pre-training on NVIDIA hardware. Pre-training — the computationally intensive phase where a model actually learns from data — remained the protected side of the line. SCMP’s coverage of the distinction was precise: LongCat-2.0 is “the industry’s first trillion-parameter model to complete full-process training and inference on a 50,000-card domestic computing power cluster.” That is a category that did not exist before June 30.

Meituan hasn’t named the chip vendor. Every announcement uses “domestic chips” or “AI ASIC superpods” without specifying a supplier or a part number. Geopolitechs notes this silence is deliberate, and the silence is almost certainly Huawei-related. Meituan used Huawei’s HCCL — the Huawei Collective Communication Library — as its NCCL replacement. AI Weekly identifies the hardware as Huawei Atlas-950 accelerators, Huawei’s commercial branding for Ascend-based products, though Meituan has not confirmed this. The HCCL dependency is the tell regardless: Huawei’s CCL is purpose-built for Huawei’s accelerators, and integrating it as your primary training communication layer is not a minor software choice. Why Meituan is staying quiet about the vendor is a question with an obvious answer given BIS’s January 2026 final rule explicitly designating Huawei Ascend chips as prohibited.

What It Took to Get Here

Running frontier-scale training on domestic chips required solving problems that NVIDIA’s ecosystem handles transparently. The solution set is now partially documented and it is not trivial. Meituan built a 6D parallelism scheme — tensor, context, expert, data, pipeline, and embedding dimensions running simultaneously — to compensate for interconnect characteristics that differ from NVIDIA’s NVLink topology. They added a 135 billion parameter n-gram embedding module, a design choice that expands the embedding space by roughly 100x using 5-grams. They implemented custom sparse attention with Streaming-aware, Cross-Layer, and Hierarchical Indexing enhancements. Technical analysis from Quasa describes the 6D approach as a direct engineering response to making training stable at this scale without CUDA native support. These are not trivial workarounds. They represent months of infrastructure work that NVIDIA customers don’t have to do.

The model circulated anonymously as “Owl Alpha” on OpenRouter for roughly two months before Meituan disclosed it — a stealth beta that generated real interaction data and validated behavior before the domestic-chip claim became a target for scrutiny. The benchmark numbers that followed disclosure are vendor-reported and not independently verified on third-party leaderboards: SWE-bench Pro at 59.5 against GPT-5.5’s 58.6, a margin FelloAI explicitly calls a statistical tie; Terminal-Bench 2.1 at 70.8; SWE-bench Multilingual at 77.3. On FORTE and BrowseComp, LongCat-2.0 trails Claude Opus 4.8. This is a model that’s competitive in agentic coding — the specific benchmark suite it targets — not one that claims frontier across all dimensions. MIT licensed, priced at $0.75 per million input tokens and $2.95 per million output, well below any Western frontier alternative. The open weights themselves were “coming soon” at the time of announcement and have since appeared on Hugging Face. The framing of this as fully open-source at disclosure was premature, though the weights now exist.

The Counter Is Real

The CFR’s position is not wrong: Huawei’s Ascend 910C delivers roughly 60% of H100 real-world performance. Huawei produces approximately 200,000 AI chips annually against NVIDIA’s millions. The CFR calculates that even under aggressive production assumptions, Huawei delivers about 5% of NVIDIA’s aggregate AI computing power by 2026, a ratio projected to decline as NVIDIA scales its output. To train LongCat-2.0 at equivalent efficiency on NVIDIA hardware would have required a fraction of the 50,000-card cluster Meituan assembled. The cost premium of domestic chips is real and measurable.

“Huawei is not a threat that justifies loosening controls; it is evidence that the controls are working,” the CFR writes. That framing has logic behind it. The controls raised costs substantially. They forced engineering workarounds that consumed years of effort. They delayed the training milestone by an estimated eighteen months to two years compared to what a comparable NVIDIA cluster would have required. AI Frontiers adds the global-infrastructure argument: outside China, AI infrastructure is overwhelmingly built on Western chips. LongCat-2.0’s competitive pricing targets customers inside China and adjacent markets, not the enterprise deals driving Western AI revenue. None of this is theater. The export controls did something.

The benchmarks are also unverified. A model that circulated anonymously for two months without being identified as a Chinese frontier system is either impressively capable or impressively selective about which tasks it demonstrated in the wild — or both. A one-point SWE-bench Pro margin needs independent reproduction to carry the strategic weight it’s being asked to carry. And CSIS documents extensively that a meaningful portion of China’s AI compute capability still runs on chips that arrived through circumvention — shell companies, smuggling, the informal channels the controls were designed to close. The domestic-chip achievement sits alongside, not instead of, a persistent black-market pipeline.

What the Binary Can’t Do Anymore

The policy argument that chip export controls meaningfully constrain Chinese AI capability rested on a specific claim: without NVIDIA hardware, frontier-scale pre-training is not possible. That claim is now false. Not because LongCat-2.0 outperforms frontier Western systems across all benchmarks — it doesn’t — but because the training run happened, under the conditions it happened, built by who built it.

Meituan is a food delivery company. Its core business is restaurant logistics and demand-matching. It is not a defense lab, a state AI research institution, or a major cloud infrastructure provider with a decade of ML platform investment. A company whose primary engineering challenge is route optimization assembled the capacity to solve 6D parallelism at scale on Huawei accelerators, build a production NCCL replacement, and run 35 trillion tokens to completion on a domestic cluster. That is not a capability that belongs to a small, controlled class of actors anymore. If Meituan can do this, the list of Chinese organizations that can do this is substantially longer than the export control framework was designed to assume.

The commercial dimension compounds the strategic one. At $0.75 input and $2.95 output per million tokens, LongCat-2.0 prices frontier-adjacent capability below what any Western provider charges. Developers in markets that have access to Meituan’s API face a specific and measurable economic choice: pay Western rates for higher benchmark scores on general tasks, or pay a fraction of that for a model that competes within a percentage point on the coding tasks that constitute most production workloads. That’s not a hypothetical pressure — it’s a real price signal in a market where inference costs are already the primary constraint on AI deployment.

Export controls were never going to function as a permanent wall. What they offered was time: time to extend compounding advantages in the chip stack and the training ecosystem, time for the US-aligned semiconductor world to build a lead that would matter at longer time horizons. CSIS assessed the controls as “only a short-term palliative for the long-term strategic challenge.” LongCat-2.0 gives that assessment a specific date.

The Ascend 910C is still 60% of an H100. Huawei’s aggregate compute is still 5% of NVIDIA’s. The 50,000-card cluster required to match results a smaller NVIDIA cluster could have produced is a real and quantifiable cost premium. None of that changes. But the argument that frontier training requires NVIDIA — and therefore that restricting NVIDIA is a meaningful constraint on Chinese AI capability — is over. The policy case for export controls now rests on narrower ground: raising the cost of each training run, slowing the rate of capability improvement, buying time for the performance gap to widen faster than domestic capability can close it. That’s a real thing to want. It just has to be argued on those terms now, not the terms that LongCat-2.0 retired on June 30.

AI-generated editorial illustration · TemperatureZero · July 5, 2026