Huawei Claims DeepSeek V4-Pro Post-Training on 1,000 Ascend 910C Chips

A research group led by Huawei Technologies claims to have completed the full-parameter post-training of DeepSeek's V4-Pro model. The team states it used a cluster of at least 1,000 Ascend 910C AI accelerator chips for the workload. This claim targets the growing debate over whether domestic Chinese silicon can handle large language model training without relying on Nvidia hardware.

Unverified assertion about tuning a trillion-parameter model without Nvidia fallback

The V4-Pro model contains 1.6 trillion parameters and requires post-training to align with instruction-following and safety guidelines. Post-training refines a pre-trained foundation model rather than building core capabilities from scratch. The Huawei-led group positions the Ascend 910C as viable for this specific tuning phase of large language model development.

The training cluster relied on 1,000 or more Ascend 910C chips to process the V4-Pro workload. Post-training focuses on adjusting a pre-trained model using instruction and safety data instead of generating foundational knowledge. The group claims the hardware handled the full-parameter tuning without requiring fallback to Nvidia accelerators.

Previous attempts to train DeepSeek models on Ascend hardware encountered unstable performance and software stack gaps. Those earlier failures forced developers to revert to Nvidia GPUs for training workloads. This new claim suggests a resolved software environment or improved chip stability, though no independent verification exists yet.

The South China Morning Post reported the Huawei-led team's assertion regarding the V4-Pro post-training run. The report carries no benchmarks, duration metrics, efficiency figures, or direct comparison to Nvidia hardware performance. DeepSeek has not commented on whether it participated in or validated the training process.

The claim rests entirely on an unverified assertion from a Huawei-affiliated research group. No public benchmarks, run durations, or efficiency data back the statement about using 1,000 Ascend 910C chips for V4-Pro post-training. DeepSeek remains silent on the matter.