Meituan Trained a 1.6T-Parameter Model With Zero NVIDIA Chips. That's the Real Headline.

TL;DR

Meituan, the Chinese company you would more likely associate with getting dinner to your door than with frontier AI, just shipped LongCat-2.0, a Mixture-of-Experts model with 1.6 trillion total parameters and about 48 billion active per token, a 1M-token context, and an MIT license. Impressive, but not the story. The story is that LongCat-2.0 was pretrained entirely on domestic AI ASICs, tens of thousands of them (reported at roughly 50,000 to 60,000 chips), with zero NVIDIA hardware anywhere in the training run. Trained on more than 35 trillion tokens across millions of accelerator-hours, reportedly with no rollbacks or irrecoverable loss spikes. This is the first time a trillion-parameter model has been pretrained start to finish off the NVIDIA stack. Weights are listed as "coming soon," a Preview is live, and the chip-sovereignty question just got a very loud answer.

Why the Hardware, Not the Model, Is the News

We have covered a lot of big MoE drops lately, from DeepSeek V4 to GLM-5.2. On raw spec sheet, LongCat-2.0 is another entry in that race: huge total parameter count, a small active slice, a million tokens of context. If that were all, this would be a footnote.

What makes it matter is the silicon. Until now, every frontier-scale pretraining run on Earth, open or closed, Western or Chinese, leaned on NVIDIA somewhere in the loop. Even DeepSeek's much-discussed domestic-chip work used home-grown accelerators for inference, while the actual pretraining still happened on NVIDIA. LongCat-2.0 closes that gap: domestic ASIC superpods for both pretraining and inference, glued together with Huawei's collective-communication library (HCCL) instead of NCCL.

Think of it like an aircraft built from entirely domestic parts. Plenty of countries can fly a foreign-engined plane. Designing and machining the engine yourself, then flying across an ocean on it without an unplanned landing, is a different class of claim. LongCat-2.0's "no rollbacks, no irrecoverable loss spikes" line is the team saying the engine held at altitude for the whole flight.

The Spec Sheet, For The Record

The model itself is a serious piece of engineering on its own terms. It is a sparse MoE: 1.6T parameters sit in the weights, but only around 48B fire for any given token. That ratio is the entire economic argument for MoE, you pay storage for the big number and compute for the small one.

Architecturally there are a couple of nods worth calling out. LongCat uses what it calls LongCat Sparse Attention (LSA) and a shortcut-layer design that lets MoE communication overlap with parallel-branch computation, which is exactly the kind of plumbing you obsess over when your interconnect is not NVLink and every microsecond of synchronization stall is multiplied across 50,000 chips. The blog also describes n-gram embeddings baked into the model, an old-school trick to give the network cheap local-context priors. None of this is revolutionary in isolation; together it reads as a team co-designing the model to fit the hardware it was stuck with, which is the whole point.

Can You Actually Run It?

Not yet, and even when you can, "running it" is doing some heavy lifting. The weights are still marked "coming soon" on the meituan-longcat Hugging Face org, with only a gated Preview live for now. When they land, the MoE math is friendlier than the headline 1.6T suggests, since only the active experts need to be resident to serve a token, but you are still looking at a model whose full weight set dwarfs anything that fits on a single box. This is a data-center model, not a homelab one. The honest summary from one early tester: if you have to ask whether you can run it, you can't.

For builders, the near-term value is not "download and serve." It is the precedent. An MIT-licensed, trillion-parameter model that you can in principle inspect, fine-tune, and self-host, produced without a single export-controlled GPU, changes the risk calculus for anyone whose supply of NVIDIA silicon is uncertain. The recipe for frontier training just got demonstrated on alternative hardware at full scale.

The Caveats, Straight

This is a Preview with a self-written announcement, and several things are not yet independently checkable. Meituan has not published a full technical report at the time of writing, the weights are not out, and there are no third-party benchmark numbers to confirm where LongCat-2.0 actually lands against the frontier. The "1.6T parameters, 35T tokens, no loss spikes" claims come from the lab, not a referee. The chip count (tens of thousands, reported around 50,000 to 60,000) is from press reporting, not a published manifest. Treat the capability question as open until the weights and a real eval pass land. The hardware claim, that this was trained off NVIDIA, is the part with the most corroboration, and it is also the part that matters most.

Key Takeaways

• The release: Meituan's LongCat-2.0, a 1.6T-parameter MoE (~48B active), 1M context, MIT license, Preview live and weights "coming soon."
• The real story: it was pretrained end to end on domestic AI ASICs, tens of thousands of them, with no NVIDIA hardware in the training loop.
• The first: earlier domestic-chip efforts (e.g. DeepSeek) used home-grown silicon for inference; LongCat-2.0 used it for pretraining too, a frontier-scale first.
• The engineering: LongCat Sparse Attention, shortcut-layer overlap, and n-gram embeddings read as a model co-designed to fit non-NVIDIA interconnects.
• The reality check: it is a data-center model, not a homelab one, and the capability numbers are still lab-reported and unverified.
• Why builders care: a viable, license-friendly frontier-training recipe that does not depend on export-controlled GPUs.

Sources: LongCat-2.0 announcement, LongCat-2.0 on GitHub, meituan-longcat on Hugging Face, South China Morning Post, Phemex News