Search
On July 6, Meituan officially announced the open-source release of LongCat-2.0, a large-scale model boasting a staggering 1.6 trillion parameters, with an average of around 48 billion active parameters. This model stands out as the first in the industry to undergo the complete training and inference process on a domestic computing power cluster composed of 50,000 cards. To overcome the constraints of memory and bandwidth inherent in domestic computing chips, Meituan implemented in-depth collaborative optimizations, spanning from model architecture design, chip adaptation, to deployment strategies. The open-source package encompasses multi-precision versions, including BF16, FP8, and INT8, along with inference code meticulously optimized for domestic computing capabilities.
-
C114 Communication Network -
Communication Home
