Huawei claims DeepSeek V4-Pro post-trained on 1,000 Ascend chips

A research team that includes Huawei Technologies has announced the successful full-parameter post-training of DeepSeek's V4-Pro, a massive 1.6-trillion-parameter model, using a cluster of at least 1,000 domestic Ascend 910C chips. The work, coordinated with the Shenzhen Loop Area Institute, Shenzhen campus of Harbin Institute of Technology, and Shenzhen Research Institute of Big Data, represents a step forward for Chinese AI accelerators which have historically struggled with training workloads due to US export controls on Nvidia hardware. The Ascend 910C is Huawei's flagship AI chip, previously shown to deliver about 60% of the inference performance of Nvidia's H100.

The post-training phase updates all model weights — contrasting with lighter fine-tuning that only adjusts adapter layers — and typically handles instruction-following, safety alignment, and task-specific data after the heavy pre-training phase. While this demonstrates that Ascend silicon can now handle a genuine training-class workload, the claim lacks key details: no benchmarks were released, no timing comparison to Nvidia hardware was given, and it's unclear how efficiently the 1,000-chip cluster was used. The announcement also doesn't prove that Ascend can handle full pre-training from scratch — the more compute-intensive and costly stage. Earlier reports indicated DeepSeek failed to complete a single training run for its R2 model on Ascend chips even with Huawei engineers present, citing unstable interconnects and software gaps. DeepSeek itself has not commented on the latest claim.