Chinese super-app company Meituan has open-sourced LongCat-2.0, a 1.6 trillion parameter artificial intelligence model. Announced on GitHub and Hugging Face, the model had been secretly tested as Owl Alpha on OpenRouter, where it dominated global rankings for two months, consuming 10.1 trillion monthly tokens. The key differentiator is its training on over 50,000 domestic Chinese ASICs, without any Nvidia GPUs.
LongCat-2.0 outperforms GPT-5.5 on SWE-bench Pro benchmark
The model uses a Mixture-of-Experts architecture with an average activation of 48 billion parameters per token and a native 1 million token context window. In standardized tests, LongCat-2.0 scored 59.5 on SWE-bench Pro, surpassing GPT-5.5's 58.6. It also achieved 70.8 on Terminal-Bench 2.1 and 77.3 on SWE-bench Multilingual, demonstrating specialization in agentic software engineering tasks. The post-training MOPD framework separates optimization into dedicated experts for reasoning, tool execution, and human alignment.
Sponsored Protocol
Aggressive pricing and MIT license to compete with Western models
Standard API pricing is $0.75 per million input tokens and $2.95 for output, but a limited-time promotion slashes costs to $0.30 and $1.20 respectively, making it one of the cheapest high-quality models. Cache hits are free, eliminating costs for repeated context. The MIT license allows commercial use without open-source obligations, encouraging enterprise adoption. The decision to train on Chinese chips marks a turning point: it proves frontier models can scale without Nvidia dependency, precisely as the U.S. restricts exports of models like GPT-5.6 and Claude Fable 5, creating a gap Meituan aims to fill, as reported by VentureBeat.
Sponsored Protocol
For more on chip manufacturing investments, read about South Korea's $518B chip fab plan. This development fits into the broader tech competition, also highlighted by the Apple Silicon roadmap leak. To learn more about Meituan, see the Wikipedia entry.
