Meituan Open-Sources 1.6T LongCat-2.0 on Chinese Chips
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Meituan released and open-sourced LongCat-2.0, a 1.6T-parameter MoE language model with ~48B active parameters and a 1M-token context window.
- The model was trained on a 50,000-card cluster of domestic Chinese processors — AI ASIC superpods, not Nvidia GPUs — according to Reuters and the company.
- LongCat-2.0 was purpose-built for agentic coding and uses a custom sparse attention architecture; observers reported it outperformed GPT-5.5 on SWE-bench Pro.
- Before formal release, the model had been silently topping OpenRouter under the alias "Owl Alpha," giving developers a real-world benchmark of its capabilities.
- Coverage from Reuters, SCMP, VentureBeat, The Next Web, and Tech in Asia converged on the framing: China's biggest AI model yet trained entirely on local chips.
- X commentator Yuchen Jin cited the release as vindicating Jensen Huang's prior point that US chip export controls won't halt Chinese AI — they instead accelerate domestic alternatives.
- Weights are being published on Hugging Face, with the model already available via APIs through TestingCatalog-listed endpoints.
Why it matters: A food delivery platform shipping a 1.6T-parameter model trained on 50,000 domestic Chinese ASICs — not Nvidia GPUs — is a concrete data point that frontier-scale AI no longer requires restricted US hardware. For US chip policy debates, this is the kind of result export controls were meant to prevent: near-frontier agentic coding performance on a fully domestic stack.
