Meituan unveils 1.6T LongCat-2.0 MoE

Training a trillion-parameter MoE on entirely domestic Chinese silicon is a massive geopolitical and technical flex from Meituan, proving that reliance on NVIDIA isn't the only path forward for massive-scale LLMs.

• –Domestic Chip Triumph: Training a 1.6T MoE model on a cluster of 50k-60k domestic AI accelerators showcases that competitive large-scale model training is viable outside of the NVIDIA/CUDA ecosystem.
• –Massive Scale, Efficient Execution: Activating 48B parameters per token out of a 1.6T total parameter pool allows the model to scale capacity while keeping inference costs relatively low.
• –Ultra-Long Context: The 1M token context window targets complex, document-heavy retrieval and reasoning tasks.
• –Release Timing: The announcement serves as a preview, with the community eagerly waiting for the weights to be uploaded to the Hugging Face repository.