Huawei Team Reports Full Post-Training of DeepSeek 1.6-Trillion-Parameter Model on Ascend 910C Hardware

*A research group that includes Huawei Technologies completed full-parameter post-training of DeepSeek V4-Pro using 1,000 domestic Ascend 910C chips.*

A Huawei-led research group states it has finished full-parameter post-training of DeepSeek's V4-Pro model. The work used 1,000 Ascend 910C chips and produced a 1.6-trillion-parameter result.

The announcement centers on post-training rather than pre-training from scratch. Post-training here refers to the stage that refines an already-initialized model with additional data and optimization passes. The group claims the entire process ran on Huawei's Ascend 910C accelerators.

The source provides no further technical details on training duration, dataset size, or measured performance gains. It also does not compare throughput or efficiency against other hardware platforms.

No independent verification or third-party benchmarks have been released. The claim rests on the group's internal report.

Why it matters

For teams that must source training hardware outside dominant Western supply chains, the result shows one path using Chinese silicon at this scale. Whether the same workload would run faster or cheaper on other accelerators remains unaddressed by the available data. The report signals continued investment in domestic alternatives but leaves open questions about reproducibility and comparative cost.

---

Sources:

• Tom's Hardware

{
  "excerpt": "A Huawei-led group completed full-parameter post-training of DeepSeek's 1.6-trillion-parameter V4-Pro model on 1,000 Ascend 910C chips.",
  "suggestedSection": "ai",
  "suggestedTags": ["huawei", "deepseek", "ascend", "model-training"],
  "imagePrompt": "Abstract rows of server racks fill a dim facility, linked by faint data pathways that suggest large-scale computation. The composition focuses on scale, connectivity, and industrial materials without devices or logos, muted color palette, cinematic lighting, 16:9"
}