The Haiguang Information Technology team successfully completed the adaptation of DeepSeek V3 and R1 models to Haiguang DCU

The Haiguang Information Technology Team recently successfully completed the adaptation of the DeepSeek V3 and R1 models to the Haiguang DCU (Depth Computing Unit), and has officially launched them. Now, users can access and download relevant models through the "Light Source" section in the "Photosynthesis Developer Community", and then quickly deploy and use these models based on the DCU platform. The DeepSeek V3 and R1 models use a number of innovative technologies, such as Multi-Head Latent Attention (MLA), DeepSeekMoE, multi-token prediction, FP8 mixed precision training, etc., which significantly improve the training efficiency and inference performance of the model.