DeepSeek V4-Pro trained on Huawei Ascend 910C chips in milestone test

Chinese researchers, in collaboration with Huawei Technologies and three Shenzhen-based research institutions, have successfully completed full-parameter post-training of DeepSeek's V4-Pro AI model using more than 1,000 Huawei Ascend 910C chips. The development, reported by the South China Morning Post, marks a significant step for China's domestic AI hardware ambitions under U.S. export restrictions. The V4-Pro model contains approximately 1.6 trillion parameters and was trained on a dataset exceeding 32 trillion tokens during pre-training—which was done on Nvidia hardware. The reported achievement focuses on post-training, which refines the model's behavior through instruction-following and safety controls, rather than the far more compute-intensive pre-training phase.

Despite the milestone, questions remain about the Ascend platform's broader training capabilities. In August, DeepSeek reportedly struggled to complete a training run for its R2 model on Ascend hardware due to unstable performance, slow chip-to-chip communication, and shortcomings in Huawei's CANN software ecosystem (its alternative to Nvidia's CUDA). As a result, DeepSeek returned to Nvidia processors for pre-training while continuing to use Ascend chips for inference. The Shenzhen announcement provided no technical details on training duration, hardware utilization, or independent benchmarks. DeepSeek has not publicly commented on the report. The news comes after DeepSeek drew global attention in early 2025 for developing competitive AI models at a fraction of the expected cost, triggering market reassessments of expensive AI infrastructure demand.