Cambricon achieves Day 0 compatibility for DeepSeek-V4 AI model enabling stable launch day operation The company used its self developed Torch MLU Ops library and vLLM inference framework to significantly accelerate model inference and improve efficiency The optimized DeepSeek V4 model features a million character ultra long context leading in capabilities like Agent performance and reasoning