Next model to be trained on domestic chips
Chinese business newspaper STAR Market Daily reported on Nov. 3 that Chinese AI company DeepSeek plans to use Chinese Huawei's "Ascend" chips to train its next large model, "DeepSeek V5." It cited several industry sources. @LeroyLi311063, an X account that posts information about China, also relayed the report on the same day.
Until now, Ascend has mainly been used for inference, running DeepSeek's models. Using it for pretraining, the most compute-intensive stage, would be a major shift in policy. The scale of training, the number of chips and the timing were not reported, and DeepSeek has made no announcement.
October report also said on 950DT by year-end
The Information reported on Oct. 19 that DeepSeek plans to begin pretraining on Huawei's new "Ascend 950DT" as early as the end of 2026. The 950DT is a chip Huawei plans to release in the October-December quarter, and much of the next "Ascend 960DT" will reportedly go to DeepSeek. It also reported that Huawei engineers are stationed at DeepSeek's offices, reflecting the model side's requests in chip design.
However, Huawei has not yet publicly shown that it can run large chip clusters stably for weeks at a time. There is also a view that DeepSeek will rely on NVIDIA equipment for large training runs for the time being.
Gradually moving to Ascend
At the end of September, DeepSeek released a compiler, computation libraries and distributed communication software for Ascend. They serve the same roles as components it has released for NVIDIA.
Huawei said in April that the DeepSeek V4 family of models can run entirely on its large computing systems. A Shenzhen research organization completed post-training of the 1.6-trillion-parameter V4-Pro on Ascend 910C, but this was not pretraining from scratch.
There is also a rumor that DeepSeek will release a large model at year-end (related article).
