Woofun AI reports that Meituan has open-sourced LongCat-2.0, a 1.6 trillion-parameter large model featuring an average activation of approximately 48 billion parameters. Designed for Agentic Coding, the architecture incorporates LongCat Sparse Attention and N-gram Embedding to optimize memory access and balance structural stability. The model underwent multi-teacher online distillation, categorizing experts into Agent, Inference, and Interaction roles, and was successfully deployed on a domestic computing cluster using the MOPD architecture.

To address constraints related to VRAM, bandwidth, and interconnects on domestic chips, Meituan implemented ScMoE for core-level parallelism and Super Kernel for reduced operator overhead. The release includes BF16, FP8, and INT8 precision versions, enabling efficient inference on existing domestic hardware through PD separation and asynchronous load balancing techniques.