| Mobile Web

Nota lightens Moonshot AI\'s Kimi K3, cutting GPU needs by up to half

AI model optimisation company Nota said on Wednesday it lightened Moonshot AI\'s open-weight model Kimi K3, reducing the number of graphics processing units needed to run it by up to half. Kimi K3 uses a Mixture of Experts architecture with 2.8 trillion parameters and typically requires eight Nvidia B300 GPUs. Nota applied its non-uniform expert pruning to remove lower-contributing modules and released two versions requiring six and four B300 GPUs.