Yesterday, Sugon's National Advanced Computing Industry Innovation Center revealed that Hygon's information technology team has finalized the localization and compatibility of the DeepSeek V3 and R1 models with Hygon's DCU, marking their official launch. These two Transformer-based models integrate two groundbreaking technologies: MLA and DeepSeek MoE. These integrations significantly reduce memory usage, enhance inference efficiency, and optimize model performance. Hygon's DCU is a high-performance GPGPU architecture AI accelerator card, which has seen widespread adoption across various sectors including science, education, finance, and healthcare. Users can now download these models from the designated platform and deploy them for use on the DCU platform.
