LongCat-2.0: how Meituan trained a 1.6-trillion-parameter model without a single Nvidia chip

A Chinese food-delivery company just released an open-weight AI model with 1.6 trillion parameters, a one-million-token context window, and a training run that used zero Nvidia hardware. I want to walk through what each of those three claims actually means, because taken together they describe a model that a year ago most people would have said could not exist yet.

The model is called LongCat-2.0. The company is Meituan. It was released on 30 June 2026 under an MIT license, which means you can use it commercially with essentially no restrictions ¹. Before the reveal, it had been quietly deployed on OpenRouter under the codename "Owl Alpha", where it had been topping the agentic-coding routing charts for weeks without anyone knowing where it came from ¹.

Three things are going on inside this release. One is architectural (how you make 1.6 trillion parameters runnable at all). One is also architectural (how you make a one-million-token context window computationally survivable). And one is infrastructural (how you train something this large on Chinese-designed ASICs instead of Nvidia GPUs). I'll take them in that order.

The parameter count is misleading, and that is the point

A dense 1.6-trillion-parameter model would be unrunnable. Not "expensive" — unrunnable. To generate a single token, a dense model has to multiply its input through every parameter it has. At 1.6T parameters in the usual half-precision format, that is roughly 3.2 terabytes of weights being touched for every token of output. No single machine has that much fast memory. The inference cost would be astronomical and the latency would be measured in seconds per token.

LongCat-2.0 is not dense. It is a Mixture-of-Experts model, usually abbreviated MoE. This is the trick that makes the 1.6T number possible.

Here is what MoE actually does. Instead of running every token through every parameter, the model is broken into a large number of smaller sub-networks called "experts". At each layer, a small routing network looks at the incoming token and picks a handful of experts to send it to. The other experts sit idle for that token. On the next token, a different handful get picked. Over a long sequence, every expert gets used — just not for every token.

The ratio that matters is total parameters versus active parameters. LongCat-2.0 has 1.6T total parameters but only around 48B active per token ¹². That is roughly a 33-to-1 sparsity ratio. The compute cost per token is closer to running a 48-billion-parameter model, which is very runnable, while the total parameter capacity of the network, the amount of stuff it can know, is closer to 1.6 trillion.

The metaphor I find most useful: imagine a hospital with 1,600 specialists but only 48 of them are on call for any given patient. The hospital as a whole knows a vast amount. The cost of any single consultation is bounded by the size of the on-call team, not the size of the hospital. MoE routing is the triage nurse deciding which specialists this particular token needs.

The metaphor breaks down in one place worth naming: in a real hospital the specialists are fixed. In an MoE model the experts are learned. During training, the router and the experts co-evolve — the router learns which experts to send which kinds of tokens to, and the experts learn to specialise in whatever the router keeps sending them. Nobody hand-designs what expert 847 does. It emerges.

One million tokens is a compute problem, not a memory problem

The second headline number is the context window: one million tokens native ¹². To put that in scale, a million tokens is roughly 750,000 words, or about ten full novels.

The reason long context is hard is not storage. Storing a million tokens is trivial. The problem is attention, the mechanism at the heart of every transformer, and it has a nasty scaling property.

Standard attention works by letting every token in the sequence look at every other token to decide what is relevant. If you have N tokens, that is N times N comparisons — quadratic in the sequence length. Doubling the context quadruples the compute. Going from 100,000 tokens to 1,000,000 tokens is not ten times more expensive — it is a hundred times more expensive.

At a million tokens, naive full attention is prohibitive. So models that claim million-token context are always doing something structural to avoid computing all N-squared pairs.

LongCat-2.0's answer is called LongCat Sparse Attention, or LSA ². The name tells you the shape of the trick: instead of attending to every previous token, each position attends to a carefully chosen subset. The subset is structured, not random, so that important long-range connections still get made, but the total number of comparisons scales much better than quadratically.

This is not a new family of ideas. Sliding-window attention (used in Mistral models) attends only to the last K tokens. Ring attention distributes attention across devices in a ring pattern. FlashAttention rearranges the computation to fit better in GPU memory. LSA is Meituan's specific pattern, tuned for their hardware and their MoE structure. The exact sparsity pattern has not been fully published as of release, which is one of the things independent researchers will want to see when the weights land.

The trade-off is real: any sparse attention pattern gives up some ability to model very-long-range dependencies compared to full attention. The engineering question is always which connections you keep and which you drop. LSA is Meituan's answer to that question. Whether it holds up on hard long-context tasks (find-the-needle-in-a-haystack retrieval, multi-document reasoning) is exactly what will get tested once the weights are public.

The hardware claim is the interesting one

Now to the part that matters geopolitically. Meituan says the entire compute stack, pre-training through inference, ran on domestic Chinese AI ASICs, not Nvidia GPUs ³. They claim roughly 50,000 chips in what they call "ASIC superpods" ³.

To understand why this is a significant claim rather than a marketing line, you need to understand what training a model this large actually requires from the hardware.

Training a 1.6-trillion-parameter MoE is not one big computation. It is thousands of chips each holding a slice of the model, all having to stay in sync. Every training step, gradients (the updates each chip has computed) have to be shared across the whole cluster and combined. This is called all-reduce, and at this scale it can dominate the total training time. If the network between your chips is slow, or the collective-communication software is immature, you will spend most of your training run waiting for chips to talk to each other rather than actually computing.

Nvidia's ecosystem has a fifteen-year head start on this problem. CUDA is the low-level programming layer. NCCL is the collective-communication library that makes all-reduce fast across thousands of GPUs. The software stack is what makes an Nvidia cluster feel like one big computer instead of ten thousand small ones.

The Chinese ASIC ecosystem, Huawei's Ascend chips, Cambricon's MLU line, and others, has been building its equivalent stack under export-control pressure since roughly 2022. Until now, the assumption in most Western analysis was that inference on domestic chips was feasible but frontier-scale training was not, mainly because the software stack for large-cluster training was not yet mature.

DeepSeek-V3, the previous marquee Chinese open-weight release, is the reference point. DeepSeek trained on Nvidia H800s and used domestic chips only for inference ³. That was a real gap. Meituan's claim is that the gap has now closed: the same 50,000-chip domestic cluster did the training end-to-end.

What would confirm the claim. Training logs showing sustained high utilisation across the ASIC cluster. Loss curves matching what you would expect from a run at this scale. A published technical report describing the collective-communication layer they used and the failure-recovery approach for a run that likely took months. As of release, none of this is public. The weights themselves are listed as "coming soon" ¹.

I want to be honest about the epistemic position here. The Owl Alpha track record on OpenRouter is real and observable — real users routed real coding tasks to that endpoint and preferred the results, for weeks, before anyone knew what it was ¹. That is a genuine performance signal on a genuine workload. It does not, on its own, verify the hardware claim. A model that trained on Nvidia hardware and lied about it would perform identically on OpenRouter. The hardware claim rests on Meituan's word until the training documentation appears.

What this actually changes

Assuming the hardware claim holds up under scrutiny — and I think it probably does, given how much reputational risk Meituan is taking on by making it — this is the first hard evidence that the training half of the export-control assumption has broken at frontier scale.

The US chip-export regime was designed around a theory: that access to Nvidia's top-tier GPUs was the bottleneck for training frontier models, and that restricting that access would slow Chinese AI development by years. Inference on domestic silicon was expected. Training was expected to remain Nvidia-dependent well into the late 2020s.

If a food-delivery company can train a 1.6T MoE on 50,000 domestic ASICs, that theory needs updating. Not abandoned — the software stack, the yield economics, and the top-of-stack performance edge all still matter — but the categorical claim that frontier training requires Nvidia hardware appears to be over.

The one thing most worth remembering from this piece: the MoE trick is what makes trillion-parameter models runnable at all, and the hardware story is what makes it possible for that trick to be done outside the Nvidia ecosystem. Both had to be true at the same time for this release to exist. That they now both are is the actual news.

Footnotes and links

Further reading

• The original Mixture-of-Experts scaling paper (Shazeer et al., "Outrageously Large Neural Networks", 2017) is still the clearest single explanation of the routing mechanism: https://arxiv.org/abs/1701.06538
• On why long-context attention is hard: Tri Dao's FlashAttention papers give the best intuition for the memory-bandwidth side of the problem. https://arxiv.org/abs/2205.14135
• On the export-control landscape: CSIS's running analysis of US semiconductor controls remains the standard reference. https://www.csis.org/analysis/updated-october-7-semiconductor-export-controls

¶