Meituan Open-Sources LongCat-2.0, a 1.6-Trillion-Parameter Model It Says Was Pre-Trained Entirely on Chinese Chips

Meituan, the Chinese on-demand-services company best known outside the country for food delivery, on June 30, 2026 open-sourced LongCat-2.0, a 1.6-trillion-parameter mixture-of-experts large language model aimed at agentic coding. The company released the weights publicly and paired the launch with a claim that lands squarely on two years of AI export-control policy: that the model completed both full pre-training and inference on a cluster of roughly 50,000 domestic Chinese AI accelerators, with no Nvidia silicon in the training loop.

If the claim holds, it is a first. Chinese labs have demonstrated large-model inference on home-grown chips for more than a year, and DeepSeek-V4-pro — the April 2026 flagship LongCat is most often compared to — used domestic accelerators for serving while heavier pre-training still leaned on established hardware. LongCat-2.0 is positioned as the first trillion-parameter model to run the entire pipeline, pre-training included, on a domestic silicon base.

LongCat-2.0 at a glance

1.6T params

A mixture-of-experts model with a context window of roughly 1 million tokens, released with open weights. Meituan says both pre-training and inference ran on a cluster of about 50,000 domestic AI ASIC superpods — the largest Chinese model it claims was trained entirely on home-grown hardware.

What Is Verifiable And What Is Not

The parameter count, the million-token context window, and the open license are the kind of facts that get checked within days as researchers download the weights and run their own evaluations. The load-bearing claim — that the entire pre-training run happened on domestic accelerators — cannot be fully verified from outside the company, and it deserves the caution any strategically convenient claim invites. Reproduction attempts, independent benchmarks, and scrutiny of the full hardware provenance (accelerators, memory, interconnect, and the manufacturing equipment behind them) will determine over the coming weeks how complete the training-decoupling story really is.

Efficiency is the open question the announcement does not answer. A run that completes is not the same as a run that completes efficiently, and it is plausible LongCat-2.0 consumed more accelerators, more power, and more wall-clock time than an equivalent run on unrestricted hardware would have. For the commercial question that matters less than it sounds; for the strategic question — whether frontier capability can be produced without the incumbent at all — it matters not at all.

Why It Matters

The significance is structural, not about any single benchmark. For three years the AI stack decoupled from the dominant hardware ecosystem from the top down — applications, open weights, fine-tuning, and inference all peeled away, in that order, because each is a comparatively fault-tolerant workload. Frontier pre-training stayed put, because keeping tens of thousands of accelerators synchronized on one job for weeks depends on a decade of mature systems software — collective-communication libraries, failure recovery, checkpointing — that lived in one ecosystem. That is the layer LongCat-2.0 claims to have crossed.

The policy implication

Friction, not absence

Export controls were built to deny training-class hardware and preserve a capability lead. A domestic cluster completing a trillion-parameter run — and the result being open-sourced worldwide on day one — suggests the controls imposed real cost and delay but did not produce absence, which was the strategic goal.

The choice to make LongCat a coding model, and to open the weights, is what gives the release reach. Agentic coding is among the most compute-hungry, highest-value workloads in the industry right now, and an open long-context model has immediate practical utility. It is the same combination — a genuinely useful model plus a free release that lets the rest of the world do the distribution — that turned DeepSeek from a domestic curiosity into a globally studied model earlier in the cycle.

For teams building developer tools, LongCat-2.0 is a new open base model to evaluate against real workloads, with the usual caveats about licensing and provenance in regulated environments. For everyone else, the takeaway is that the world now has two independent silicon bases capable of producing frontier-scale models where it recently had one — a reduction in concentration risk regardless of whether you ever touch a Chinese chip.

This is the reported edge of a larger shift. For the full argument on what the training decoupling means — why training was the last lock, how a delivery company ended up breaking it, and what the incumbent hardware ecosystem should do now — see the companion analysis, The Training Decoupling: China Pre-Trained a Frontier Model Without Nvidia. It sits alongside the inference-silicon turn as the two halves of the same story: the serving layer fragmented first, and now the training layer has a second home.

Sources

Meituan LongCat-2.0 open-weight release and technical notes (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"
Cryptobriefing / AI Market Watch coverage of the 1.6-trillion-parameter release and 50,000-card domestic cluster