Huawei cluster post-trains DeepSeek V4-Pro without Nvidia hardware

A Huawei-linked Shenzhen research group says it completed full-parameter post-training of DeepSeek V4-Pro, a 1.6-trillion-parameter model, using a cluster of at least 1,000 domestic Ascend 910C accelerators. Full-parameter post-training updates every model weight after pre-training, making this a more demanding demonstration than inference or a lightweight adapter tune. Tom's Hardware, citing the Shenzhen municipal government and South China Morning Post, says the work involved Huawei, the Shenzhen Loop Area Institute, Harbin Institute of Technology's Shenzhen campus, and the Shenzhen Research Institute of Big Data. The result does not prove that Chinese chips can yet match Nvidia for frontier pre-training, and it remains a reported team claim rather than an independently reproduced benchmark. It does show that export controls are accelerating a credible domestic training stack for large-model refinement.

Key details: June 6, 2026, DeepSeek V4-Pro, 1.6 trillion parameters, At least 1,000 Ascend 910C chips, Full-parameter post-training.

Continue swiping for more AI Brief stories.