Meituan, the Chinese food-delivery giant, has open-sourced LongCat-2.0, a 1.6-trillion-parameter MoE model with a claimed million-token context window that finished its entire pre-training and inference run on a 50,000-chip domestic cluster. The company positions the model for agentic coding and repository-scale work, saying results sit near Gemini 3.1 Pro or DeepSeek V4-pro levels, yet the announcement immediately drew pushback from forecaster Peter Wildeford over whether the hardware performance gap still leaves China years behind.
Meituan, 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 of AI that runs on Chinese chips.
Wildeford’s core point is that raw chip count tells only half the story; per-chip throughput and interconnect quality still trail leading Western stacks by a meaningful margin. Without independent benchmarks that isolate those variables, it stays unclear how much the 50,000-chip cluster actually narrows the capability distance.
The model ships under an MIT license on GitHub and Hugging Face, with full weights promised shortly. Early anonymous testing under the name Owl Alpha reportedly topped OpenRouter leaderboards for developer tasks, giving practitioners an immediate, no-cost way to test its agentic coding claims against their own repos.
Many users questioned Meituan's 1.6T parameter model trained on Chinese chips as hype or exaggeration while others praised the efficiency gains and necessity-driven progress under export controls.
No Digg Deeper questions have been answered for this story yet.
@Yuchenj_UW Checked with friends at Meituan, and it's confirmed that the chips they are currently using are all Chinese chips, and the performance is quite impressive.
«I find it hilarious that people think training an AI model on chips is an accomplishment... that's what chips are supposed to do! obviously…» First principles thinker, superforecaster in all his glory
I find it hilarious that people think training an AI model on chips is an accomplishment... that's what chips are supposed to do!
obviously what matters is (a) how many chips you have and (b) how good the chips are -- China is still very behind on both
@Yuchenj_UW @cyodyssey 听说昇腾950DT提前上市了，不知道是不是这个
@Yuchenj_UW china was going to figure it out anyway sooner or later. the key is that the US has to remain competitive by encouraging more innovation.
that's the american spirit
@cyodyssey Huawei? 👀
@ColinZhouUSA
@JidanzaiDashan @LeZhengX @Yuchenj_UW @cyodyssey 华为昇腾 Ascend 910B 单卡 HBM2e 内存 64GB，显著小于80GB。昇腾服务器节点常标配双端口 200G RoCE 网卡（或多口 200G），支持高性能 RDMA 规模互联。这与博客描述完全匹配，很可能是 Meituan LongCat-2.0 训练用的 5-6 万张国产芯片主力型号。其他 Cambricon 等也在类似栈上优化。
Worth adding: this model was already running anonymously on OpenRouter as ‘Owl Alpha’ for two months before Meituan revealed it was LongCat-2.0 — and it was topping developer usage charts the whole time, nobody knew it was Chinese-chip-trained. The capability gap closed quietly before the export control story even broke.
@Yuchenj_UW Which chips specifically? This is so vague
@Yuchenj_UW china was always going to roll their own chips. the evidence is that they roll their own version of everything ever. the idea that the now legal presence of nvidia chips would stop or meaningfully slow this is contrary to examples from numerous other industries.
@Yuchenj_UW Training at that scale on 50K chips proves necessity truly speeds up the leapfrog.
@LeZhengX @Yuchenj_UW @cyodyssey 博客里说的单卡芯片显著小于80GB，自带200G网卡。@grok 有哪些中国芯片满足这些指标？
@Yuchenj_UW export controls can work for some cases, but not for china. huawei, moore threads are doing some insane work there.
@Yuchenj_UW I don’t even think it’ll accelerate it that much, they are already doing everything they can to get technological independence. Similar to how US is already working on rare earth independence, it’s just a long process but it will eventually happen
@Yuchenj_UW Do you actually believe that lol. If they did, then you might want to short Nvidia
@Yuchenj_UW The only issue is the amount of energy used, but china is already prepared for this, so it's more a question of whose LLM will be used more and take out the competition here. Way too many AI companies in China already.
@Yuchenj_UW Training a 1.6T model on 50K homegrown chips is wild, but making things actually run efficiently—especially with CUDA’s years of tooling—still matters way more than just stacking up scale.
@Yuchenj_UW The chips only matter for neural networks an LLMs. Eventually common conventional CPUs will run much more sophisticated AIs.
AI that's more than just autocomplete... it's coming.
@teortaxesTex if it can run inference it can run trainining but the okane it's on selling tokens.
@__vandos__ @Yuchenj_UW Thanks for sharing - I am guessing it is this link you are referring to: https://openrouter.ai/openrouter/owl-alpha ?
How can we know it is topping the developer usage chart?
Also, the prompt tokens is much higher than completion tokens, isn't that a signal that it is not developer usage?
