Huawei said DeepSeek V4-Pro and V4-Flash have been officially released and
open-sourced on Apr 24, with full support across its Ascend computing stack.
DeepSeek’s latest models are optimized for deployment on Ascend supernode
systems, with Ascend 950 and A3 series supporting inference and training. Huawei
said performance improvements come from kernel fusion, multi-stream parallelism
and quantization techniques, enabling higher throughput and lower latency.