Meituan's LongCat team releases 1.6T-parameter AI model trained on domestic ASICs

Why it matters

LongCat-2.0 turns Meituan's AI work into a test of whether frontier-scale training can move off Nvidia-dependent infrastructure without falling out of developer workflows.

Meituan LongCat (@Meituan_LongCat) released LongCat-2.0 on June 30, positioning the 1.6 trillion-parameter mixture-of-experts model as proof that a Chinese internet company can train and serve a frontier-scale system on domestic AI ASIC infrastructure rather than the Nvidia stack that still defines most large-model economics.

The headline number is scale: Meituan says LongCat-2.0 has 1.6 trillion total parameters and about 48 billion parameters activated per token. The model is designed for coding and agentic workloads, with Meituan saying it trained on more than 35 trillion tokens and used hundreds of billions of tokens of 1 million-context data. The Hugging Face model card and GitHub repo repeat the same core claims, while the GitHub repository is marked MIT-licensed.

That makes the release consequential, but not yet cleanly open in the way developers usually mean it. The public repo and Hugging Face page both say the model weights are "coming soon." Meituan is calling LongCat-2.0 an open-source release, and the repository is public, but the operationally important artifact - the full weights - was not available from the model page at launch. For developers, that distinction matters: an MIT repo with a model card is a signal; downloadable weights are the product.

The sharper point is not that Meituan, best known as a local-services and delivery platform, has an AI lab. It is that Wang Xing's company is using the LongCat line to attack two bottlenecks at once: model capability and compute supply. South China Morning Post reported that Meituan claims LongCat-2.0 is China's largest AI model trained entirely on home-grown hardware and that it completed both pre-training and inference on a 50,000-card domestic compute cluster. Meituan's own technical note is more specific on the system design: LongCat-2.0 was pre-trained on more than 50,000 AI ASICs, and the company says the full training run and deployment were built on large-scale AI ASIC superpods.

That is the strategic claim underneath the release. US export controls have made access to advanced Nvidia accelerators a structural constraint for Chinese AI labs. Meituan is arguing, through a model release rather than a policy memo, that the constraint can be engineered around. The company says it had to build deterministic operators, fault recovery, memory management, long-context training infrastructure and superpod scheduling because the non-Nvidia software ecosystem is less mature. In other words: the model is also a public benchmark of the stack beneath it.

The architecture bet

LongCat-2.0 builds on the earlier LongCat-Flash line, which OpenRouter lists as a 560 billion-parameter MoE model with roughly 27 billion average active parameters. LongCat-2.0 nearly triples the total parameter count while keeping active parameters far below the headline size, the basic economic promise of MoE systems: buy scale in total capacity without paying dense-model inference cost on every token.

Meituan says the new model adds LongCat Sparse Attention, or LSA, for long-context workloads. The company describes LSA as an evolution of DeepSeek Sparse Attention with a lighter indexer intended to reduce the cost of ultra-long input processing. The release also adds a 135 billion-parameter N-gram Embedding component, which Meituan says expands the embedding space by roughly 100x through token combinations while keeping that component under 10% of the total parameter budget.

Those are company claims, not independent audit results. But they show where Meituan is aiming: repository-scale coding, long-running agent tasks and tool-heavy workflows where long context and stable serving matter more than chatbot polish. The company says LongCat-2.0 is integrated with Claude Code, OpenClaw and Hermes-style harnesses for code understanding, repository edits and automated task execution.

The OpenRouter shadow launch

The release also confirms what developers had been circling for weeks: LongCat-2.0-Preview appears to have been running in the market before the official reveal. The user-supplied source linked an X post that is no longer accessible, but the claim it surfaced is consistent with outside reporting that the OpenRouter stealth model "Owl Alpha" was LongCat-2.0-Preview.

The OpenRouter usage number is harder to state as a clean fact because OpenRouter's public rankings page describes live usage and task rankings but does not expose a stable static token table in the retrieved page. Third-party reports put Owl Alpha at roughly 10 trillion to 11 trillion monthly tokens. Crypto Briefing reported 10.1 trillion monthly tokens and said the model was LongCat-2.0-Preview; KuCoin carried a flash item citing 11 trillion monthly token throughput. Those are not OpenRouter confirmations, but they point to the same pattern: LongCat gained meaningful developer usage before Meituan put its name on the full release.

That matters because distribution is increasingly part of model quality. A lab can publish benchmark tables, but agent developers vote with routing layers, free quotas and latency. If LongCat-2.0-Preview reached that volume through OpenRouter before the brand reveal, Meituan did not just ship a large model. It got developers to test it in workflows where switching costs are low and performance complaints surface quickly.

The open question

The release leaves two gaps. First, the full weights are not yet live on the model page, so developers cannot fully evaluate the open-source claim on their own hardware. Second, Meituan has not named the domestic ASIC suppliers behind the training run in the public materials reviewed here. That omission is not cosmetic. If LongCat-2.0 is meant to prove that Chinese frontier-scale training can move beyond Nvidia, the specific hardware supply chain is part of the story.

Still, the direction is clear. Meituan is using LongCat to turn a consumer internet balance sheet into AI infrastructure leverage. The company that built its moat around dense urban logistics is now applying the same operating logic to compute: route around bottlenecks, squeeze throughput from constrained infrastructure, and put the product where developers already are. LongCat-2.0 is not just another Chinese open model. It is Meituan's argument that the next AI platform fight will be won as much in systems engineering as in model weights.