Meituan Open Sources LongCat-2.0: 1.6 Trillion Parameter Model Trained on Chinese Chips

Chinese delivery app company Meituan has officially open-sourced LongCat-2.0 on GitHub, Hugging Face, and its native platform, revealing it as the engine behind the anonymous 'Owl Alpha' model that topped OpenRouter charts for two months. The 1.6-trillion-parameter Mixture-of-Experts (MoE) system features a native 1-million-token context window and is released under a permissive MIT license for commercial use.

The pricing strategy is aggressive: all context-cache hits are free, with a limited-time promotion reducing uncached input to $0.30 per million tokens and output to $1.20 per million tokens. Standard pricing is $0.75/$2.95 per million tokens in/out, competitive with top global models. This move aims to disrupt closed-source enterprise dominance in autonomous software engineering.

A key differentiator is that LongCat-2.0 was trained entirely on over 50,000 domestic Chinese ASICs, proving that near-frontier AI models can be scaled without relying on US Nvidia GPUs. This successful deployment of alternative silicon signals a structural shift that could threaten Nvidia's dominance in AI training hardware.

The release comes amid US pressure on American labs to restrict access to their latest models, such as OpenAI's GPT-5.6 and Anthropic's Claude Fable 5/Mythos 5. Experts warn these defensive regulations have backfired, driving developers toward affordable open-source alternatives like LongCat-2.0. The model's unbranded residency on OpenRouter processed 10.1 trillion monthly tokens, a 242% month-over-month surge.

Technically, LongCat-2.0 employs aggressive MoE sparsity, activating only 33-56 billion parameters per token out of 1.6 trillion total. It introduces LongCat Sparse Attention (LSA) to efficiently handle the 1-million-token context window, overcoming quadratic scoring costs and memory fragmentation through streaming-aware indexing and other innovations.