China Trains a Frontier-Scale Model Without Nvidia: Inside Meituan's LongCat-2.0

A food-delivery company just moved the export-control debate

The most consequential model release of the past week did not come from OpenAI, Google, or Anthropic. It came from Meituan, the Chinese food-delivery and local-services giant, which on June 30 open-sourced LongCat-2.0, a 1.6-trillion-parameter language model. The headline number is not the parameter count — plenty of labs have crossed the trillion-parameter line — but the hardware underneath it. Meituan says LongCat-2.0 was pre-trained and served entirely on a cluster of more than 50,000 domestically made AI ASICs, with no Nvidia silicon in the loop.

If that claim holds up, it is the clearest evidence yet that the central bet behind US export controls — that denying China the latest Nvidia stack would keep it from training frontier-adjacent systems at scale — is weaker than Washington assumed. That is why a model from a company better known for restaurant logistics is being discussed alongside GPT-5.5 and Claude Opus this week.

What was actually released

Per SCMP, LongCat-2.0 is a 1.6-trillion-parameter model with a 1-million-token context window, open-sourced under a permissive license (reported by several outlets as MIT). It is a sparse Mixture-of-Experts design, meaning only a fraction of those 1.6 trillion parameters — secondary reporting puts active parameters in the tens of billions per token — actually fire on any given request, which keeps inference cost far below what the headline size implies. The Decoder adds that the model was trained on more than 35 trillion tokens.

The distinction Meituan is pressing is where the training happened. China's previous flagship achievements — including DeepSeek's V4-pro — leaned on domestic chips mainly for inference, the lighter task of answering queries, while the far more demanding pre-training stage still depended on Nvidia hardware. Meituan's claim is that LongCat-2.0 ran the full pipeline, pre-training included, on domestic ASICs. That is the part that matters, and the part that is hardest to independently confirm.

The benchmark picture: strong in code, still trailing at the frontier

On the numbers Meituan has published, LongCat-2.0 is genuinely competitive on agentic coding and noticeably behind on hard reasoning. The Decoder reports scores of 59.5 on SWE-bench Pro and 77.3 on SWE-bench Multilingual, where the model edges out Gemini 3.1 Pro and GPT-5.5. Those are the results driving the "near-frontier" framing.

But the same reporting is careful about the ceiling. On reasoning- and instruction-heavy tests — IFEval (90.0), IMO-AnswerBench (81.8), and GPQA-diamond (88.9) — LongCat-2.0 trails the leading Western systems, and it falls short of Claude Opus 4.7 and 4.8 even on the coding benchmarks where it otherwise shines. The honest read is a model that is excellent at a specific, commercially valuable slice of work (writing and fixing code) and merely good elsewhere. That profile is not an accident: coding is where open-weight Chinese models have concentrated, because it is measurable, in demand, and less dependent on the polished instruction-following that closed labs spend heavily to tune.

One important caveat on all of this: The Decoder notes the model was not yet on Hugging Face at the time of writing, "making independent verification difficult." The benchmark figures are, for now, Meituan's own. Several aggregators also report that the model had quietly topped OpenRouter usage under the codename "Owl Alpha" before its identity was revealed — a striking detail, but one I could not confirm in the primary reporting, so treat it as unverified.

Why the hardware claim is the real story

The chips are where skepticism should focus, and also where the significance lives. SCMP describes the training as running on "large-scale clusters of tens of thousands of AI ASIC superpods" — application-specific chips built for a narrow workload rather than general-purpose GPUs. Crucially, neither SCMP nor The Decoder names the chip vendor; The Decoder states plainly that "Meituan didn't name the specific chip maker." Secondary coverage points toward Huawei's accelerator line and its HCCL communication library (the domestic analogue to Nvidia's NCCL for coordinating thousands of chips), but that attribution is not confirmed in the primary sources.

Why would Meituan stay vague? Naming a supplier invites scrutiny — of yields, of real-world cluster utilization, and of how much Nvidia-trained tooling and data quietly shaped the result. Training a 1.6T model end-to-end on 50,000 non-Nvidia accelerators is as much a distributed-systems and networking feat as a modeling one; interconnect and communication libraries are exactly where domestic stacks have historically struggled. If Meituan genuinely solved that at this scale, the vagueness may be strategic rather than evasive.

Hype versus what's confirmed

Strip it down and three things are well-supported: the model exists and is open-weight; it posts strong self-reported coding benchmarks; and Meituan claims a fully domestic training pipeline at roughly 50,000-chip scale. Three things are not yet confirmed: independent reproduction of the benchmarks, the identity and true performance of the chips, and how efficiently that cluster actually ran versus an Nvidia baseline.

The gap between "trained without Nvidia" and "trained as efficiently as with Nvidia" is where the policy stakes sit. Export controls were never meant to make Chinese frontier training impossible — only slow, expensive, and inefficient. LongCat-2.0 doesn't refute that logic; a model can be a real achievement and still have cost far more compute-hours than a Western equivalent. What it does puncture is the stronger, load-bearing assumption that denying the newest Nvidia stack creates a hard ceiling. The ceiling, this release suggests, is higher than advertised — even if the climb is harder.

The takeaway

LongCat-2.0 is the most important model story of the week not because it beats the frontier — it doesn't, cleanly — but because of who built it and on what. A logistics company shipping a 1.6-trillion-parameter, MIT-adjacent open model it says was pre-trained end-to-end on domestic ASICs is a data point Washington's chip strategy has to reckon with. The right posture is measured: the benchmarks are self-reported and unreproduced, the chips are unnamed, and efficiency claims are untested. But the direction is unmistakable. Each release like this narrows the window in which export controls buy meaningful time, and shifts the question from whether China can train frontier-scale models on its own silicon to how efficiently — and how soon the answer stops embarrassing anyone. Watch for the Hugging Face weights and independent benchmarks; that's when hype either hardens into fact or quietly deflates.