Meituan has announced LongCat-2.0, a 1.6 trillion parameter model trained entirely on domestic Chinese hardware without any Nvidia components. The team achieved this using a cluster of over 50,000 locally made AI ASICs to process more than 35 trillion tokens. Despite forming only in 2023, the group has already released two versions, with the latest demonstrating that large-scale training is possible within China’s current export restrictions.
The results show mixed performance compared to Western rivals. LongCat-2.0 outperforms Gemini 3.1 Pro and GPT-5.5 on specific coding and multilingual tasks but trails significantly behind Claude Opus 4.7 and 4.8 on general reasoning benchmarks. This development highlights a shift in capability rather than a complete replacement for foreign chips. Independent verification remains difficult because the model has not yet been published on HuggingFace.
- Training relied on over 50,000 domestic AI ASICs
- Model processes more than 35 trillion tokens
- Specific chip manufacturer names were not disclosed
