AI model lightweighting and optimisation company Nota said on Wednesday it has lightened Moonshot AI's AI model Kimi K3, cutting by up to half the graphics processing units (GPUs) needed to run it.

Kimi K3 is an open-weight AI model built on a Mixture of Experts (MoE) architecture with 2.8 trillion parameters. It selectively activates only the expert modules needed based on the input to improve computing efficiency, but it requires eight Nvidia B300 GPUs to run.

Nota explained that the model is designed from the training stage with weights and activations reduced to 4-bit and 8-bit, respectively, limiting additional compression using conventional quantisation methods.

Nota said it lightened the model by applying its self-developed Non-uniform Expert Pruning technology. The technique analyses layer-by-layer characteristics and module importance, preserving key experts while selectively removing modules with low contribution.

Nota unveiled 2 lightened versions that reduce expert modules by 25 percent and 50 percent, respectively. As a result, the B300 GPUs needed to run the model fell from 8 to 6 and 4, respectively.

In its own evaluation, Nota said the 50 percent lightened model posted a higher performance retention rate across multiple benchmarks compared with REAP (Router-weighted Expert Activation Pruning), which removes layer-by-layer experts at the same ratio.

Kim Tae-ho (김태호), Nota's chief technology officer, said, "We will expand the scope of technology application so that various frontier-level AI models can be used efficiently even with limited computing resources."