Meituan Releases LongCat 2.0 AI Model Trained on Domestic Chinese Chips
Meituan has released LongCat 2.0, a large language model with 1.6 trillion parameters trained entirely on a cluster of domestic Chinese AI chips. The model utilizes a mixture-of-experts architecture that activates approximately 48 billion parameters per inference, supports a 1 million token context window, and was trained on 35 trillion tokens. It currently powers the Owl Alpha model available for testing on OpenRouter and will be made publicly available on Hugging Face shortly.
The release highlights how Chinese technology companies are advancing artificial intelligence despite U.S. export restrictions limiting access to advanced Nvidia graphics processing units. Meituan, China’s largest food delivery platform, trained the model on approximately 50,000 domestic AI chips, reportedly hardware from local manufacturers such as Huawei. To compensate for the lower individual performance of domestic accelerators, the company engineered custom parallelism methods, sparse attention mechanisms combining techniques from existing frontier models, and deterministic operators to ensure a stable training run without loss spikes. Industry analysts note the achievement underscores a broader shift toward software and architectural optimization to overcome hardware compute scarcity.
From the sources (25 posts)
@teortaxestexWHOA Ok, here it goes. Owl Alpha. V4-scale model, updated DSA, N-gram Embeddings, decent evals, «pre-trained on over 50K AI ASICs … superpods» from *checks notes* Meituan. Yeeeah. these are clearly *not* Ascend 950DT SuperPODs (<<80 G
@teortaxestextbh as usual very interesting technical details from Meituan but when I click "try it" I get something like Qwen-4B, seemingly worse than their first LongCat-Flash. Completely fails in Russian too. Idk if that's their new model https://t.co
@teortaxestexDamn, that's right. So: "≈48B" (thanks to N-Gram embedding, variable) active, 35T tokens. V4-tier, ≈8e24? Would be the biggest Chinese pretraining on domestic hardware. Some strange "Superpods": > "our accelerators" > "up to 48 machines e
@emostaqueMost popular model on @OpenRouter (10tr tokens) turns out to be a 1.6tr MoE by @Meituan_LongCat (superapp/DoorDash of China) Basically Gemini / Opus 4.6 level 35tr tokens trained entirely on 50k Chinese ASICs No GPUs needed https://t.co
@teortaxestexHm, yes, it does match if you remember that 910C is two "910B" dies glued together. Okay then! 384 CloudMatrix – or rather, a hundred or so of them – is good enough to train a 1.6T MoE, it seems. This is 2025 tech. In ≈3 weeks we'll see ac
@teortaxestexEven I didn't notice this. Btw Tencent invested in Enflame (though they invest in everyone). One way or another, Meituan now trains large models on domestic compute. EDIT: Enflame is likely irrelevant
@eliebakouchnew 1.6T total, 48B active open model coming very soon from @Meituan_LongCat custom sparse attention, n-gram, adaptive expert computation, 1M context, 35T of training tokens, muon optimizer, multi teacher OPD, trained on 50k chinese gpus
@bigeagle_xdRT @yzhang_cs: I've read these two LongCat reports many times. The MoE design and the ngram embed ablations are quite impressive. https://t…
@yuchenj_uwMeituan, basically China’s DoorDash, trained a 1.6T parameter LLM on 50K Chinese chips. It reminds me of Jensen Huang’s point on the Dwarkesh podcast: export controls on Nvidia GPUs won’t stop China. They’ll just accelerate the development
@eliebakouchthe new sparse attention method introduced with this model is basically a combination of components from existing ones. let's go over each sparse attention method and what they keep from them: - deepseek sparse attention (DSA): they keep t
@poezhao0605Meituan, China's largest food delivery platform, open-sourced a 1.6 trillion parameter AI model today. LongCat-2.0 was trained from scratch on a 50,000-card domestic GPU cluster. Native 1M context. 30T+ tokens of pretraining data. A deliv
@poezhao0605The architecture follows the MoE efficiency playbook that DeepSeek popularized: 1.6T total parameters, roughly 48B activated per inference call (dynamic range: 33B to 56B). What stands out is not the spec sheet. What stands out is who buil
@poezhao0605The structural signal: Meituan claims this is the first trillion-parameter model trained end-to-end on a domestic chip cluster. That claim needs independent verification. The direction does not. Chinese labs are not waiting for chip parity
@lateinteractionwhy aren’t retrieval people working on sparse attention, all the techniques look like poor man’s IR
@eliebakouchle chaton long > le chaton fat for now
@mweinbachThis model was trained entirely on ASICs from China, inference is entirely on ASICs from China It's not US frontier level, but it's incredibly competitive and pretty damn close It was trained on 35T tokens with no rollbacks or irrecoverab
@_akhaliqLongCat-2.0 dropping on Hugging Face soon
@tengyanaithey hand-built the parallelism and deterministic operators and pushed a 35T-token run with no loss spikes. bridging the software/systems gap on immature hardware watch out for China
@andrew_n_carrRT @Meituan_LongCat: Introducing LongCat-2.0 🐱 1.6T parameters · MoE with ~48B active · 1M context The full model behind Owl Alpha on @Open…
@mweinbach@franktheglock I don't even care about model performance tbh It's that they trained the model and it's a stable run, and it was done entirely on domestic ASICs at a 1.6T size and 1M context window that's essentially China frontier archite
@mervenoyannI was just mentioning today how counterintuitive it is because it pushes Chinese labs to be super creative around architecting models
@matvellosoNecessity is the mother of invention
@teortaxestexRT @eliebakouch: the new sparse attention method introduced with this model is basically a combination of components from existing ones. le…
@teortaxestexFirst giant Chinese LLM pretrained on domestic NPUs. First serious Japanese pretrained LLM in a long while iirc (Fugu doesn't count it's an orchestrator) The East is Rising…
@quixiaiRT @Meituan_LongCat: Introducing LongCat-2.0 🐱 1.6T parameters · MoE with ~48B active · 1M context The full model behind Owl Alpha on @Open…