Meituan open-sources 1.6T model on 50K Chinese chips
SkimNews Take
Reusing existing models while building out domain-specific tool integrations suggests the differentiating layer is shifting from the model itself to the curated ecosystem of data connections and workflows wrapped around it.
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Meituan (3690.HK) released and open-sourced LongCat-2.0, a 1.6T-parameter MoE language model with ~48B active parameters, 1M-token context, and 30T+ tokens of pretraining data
- The model was trained entirely on a 50,000-chip cluster of domestic Chinese ASICs, with no Nvidia GPUs in the stack
- LongCat-2.0 was already leading OpenRouter under the codename 'Owl Alpha' before its official release, with VentureBeat calling it a 'near-frontier agentic coding model'
- The architecture introduces LongCat Sparse Attention (LSA), designed for efficient 1M-context inference on agentic coding workloads
- Felix on X claims LongCat-2.0 outperforms GPT-5.5 on SWE-bench Pro, though the benchmark comparison is not independently verified in the source
- Meituan is known primarily as China's largest food delivery platform—not a frontier AI lab—making the training run notable for who built it, not just what was built
- Observers including Yuchen Jin framed the release as confirmation of Jensen Huang's own warning that US GPU export controls 'won't stop China' but will 'accelerate the development of AI that runs on Chinese chips'
Why it matters: A Chinese food delivery company (Meituan, 3690.HK) trained a 1.6T-parameter model on 50,000 domestic ASICs that was already topping OpenRouter before its official release. As Yuchen Jin noted, this echoes Jensen Huang's own admission that export controls accelerate rather than block domestic Chinese chip development—undermining the strategic premise of Washington's GPU restrictions by demonstrating a non-tech firm can scale frontier-class models on homegrown silicon.
