DeepSeek plans to train its next flagship model, V5, on Huawei’s Ascend processors, according to STAR Market Daily, which cites several industry sources. Geopolitechs relayed the report. Neither company is quoted, and nothing has been independently confirmed.
So far Ascend chips have mostly run DeepSeek’s finished models for users. Teaching a model from scratch, known as pre-training, is far heavier work. A Shenzhen research consortium recently finished post-training V4-Pro (a model with 1.6 trillion parameters) on a cluster of Ascend 910C chips, but Geopolitechs points out that this is a lesser job.
The groundwork is visible. Huawei says its Ascend SuperPoD systems can fully support the V4 family, and DeepSeek released compiler tools, computing libraries, and communication software for Ascend as open source on September 30.
If V5 really trains on Ascend, Chinese developers gain proof they can build frontier models with less dependence on Nvidia, and Huawei gains a showcase customer. A stalled or unstable run would show the opposite. Geopolitechs names training stability, cluster performance, and software maturity as the open questions.
Until DeepSeek or Huawei confirms it, treat V5 on Ascend as a plan, and judge it by whether V5 ships.
Reported by Geopolitechs, citing STAR Market Daily and unnamed industry sources.
