News
DeepSeek V4 Runs on Huawei Ascend 950 — First Trillion-Parameter MoE Deployed on Pure Domestic Chinese Chips
Huawei confirmed its Ascend 950-based supernode clusters support DeepSeek V4, marking the first trillion-parameter MoE model deployed without any NVIDIA hardware. V4's architecture shifted from CUDA to Huawei's CANN framework, with Cambricon and Moore Threads chips also accelerating synchronously. This represents a significant milestone in China's push for compute self-sufficiency under US export controls.
Source
↳ Follow the thread