DeepSeek-V4 Large Model Launches, Cambricon Achieves Day 0 Adaptation and Opens Source Code

The domestic large model DeepSeek-V4 was launched today, including the 285B-parameter flash version and the 1.6T-parameter pro version. Cambricon completed Day 0 adaptation based on the vLLM inference framework immediately, with the adaptation code open-sourced on GitHub.

This adaptation relies on Cambricon's NeuWare software ecosystem, collaboration with Zhongzhi FlagOS, and self-developed CNAgent. It enables fast migration. Meanwhile, optimizations like Torch-MLU-Ops fusion operator library and vLLM's 5D hybrid parallel release the model's inference potential, reduce computing costs, and provide efficient deployment solutions for China's AI industry.