- Meituan released LongCat-2.0, a 1.6 trillion parameter open-source LLM, trained end-to-end on domestic Chinese hardware without NVIDIA chips.
- It's the industry's first trillion-parameter model to complete both training and inference on domestic hardware, marking a major milestone for China's tech self-reliance.
- The model uses Huawei's Atlas-950 SuperPoDs and HCCL software, mirroring NVIDIA's setup, and shows strong performance but still trails top global models like GPT-5.5.
- The launch demonstrates frontier-scale training is viable on Chinese hardware, potentially narrowing the gap with Western systems faster than expected.
- Challenges remain, including a less mature software ecosystem and memory limits compared to NVIDIA's banned H800 chips.