China's Meituan Open-Sources 1.6T LongCat-2.0 — Trained Entirely on Domestic Chips
Summary: Meituan open-sourced LongCat-2.0, a 1.6-trillion-parameter model trained entirely on Chinese ASIC clusters, delivering near-frontier agentic coding performance at a fraction of US rival costs — and publicly proving that export controls haven't stopped China from scaling frontier AI.
Key Facts
- Scale: 1.6T parameters, 1M-token context window, released under MIT license
- Hardware: pre-trained on a cluster of 50,000+ domestic Chinese ASICs — first confirmed case of both pre-training and inference on entirely Chinese-made silicon (DeepSeek-V4 used domestic chips only for inference)
- Performance: held the top position on OpenRouter's coding leaderboard for several months, competing with GPT-5.5 and Gemini 3.1 Pro
- Cost advantage: reportedly runs at 5–10% of GPT-5.5's API cost per token at equivalent parameter scale
- Distribution: weights publicly available; API accessible via Meituan Cloud and third-party partners
Why It Matters
This is the clearest public evidence yet that US export controls have not created a fundamental ceiling on Chinese frontier AI development. Completing the full training cycle — pre-training, not just inference — without Nvidia hardware removes the most commonly cited bottleneck. Policymakers debating chip restrictions and AI labs tracking competitive dynamics will treat LongCat-2.0 as a significant data point regardless of how its benchmark claims are independently verified.
Read More
- VentureBeat — VentureBeat
- SiliconANGLE — SiliconANGLE
- Cryptopolitan — Cryptopolitan