A Shenzhen Hetao College AI training-platform team, with Harbin Institute of
Technology (Shenzhen), Shenzhen Institute of Big Data and Huawei groups, using
the Shenzhen Zhicheng AI compute platform and a domestic Ascend 910C AI compute
cluster, completed full-parameter training of the 1.6 trillion-parameter model
DeepSeek-V4-Pro. The exercise is a rare instance globally of a third-party team
completing training at this scale on a domestic compute platform, yielding
operational experience and supporting the claim that domestic AI chips can
sustain world-class, very large-parameter model training. The team emphasized
that running inference on existing models and conducting from-scratch
full-parameter training are distinct technical stages with materially different
hardware requirements; the results indicate domestic AI compute can handle
top-tier training workloads and that the technical path is feasible.
