What's this about?
A systems report on full-parameter post-training of the trillion-parameter DeepSeek-V4 MoE family on a Huawei Ascend SuperPOD — memory pressure, communication overlap, and the engineering to make it work at scale.
What exactly is "post-training"?
The phase after pre-training — RLHF, reasoning RL, fine-tuning — that turns a raw text predictor into a useful assistant. Doing it "full-parameter" at trillion scale (rather than cheap adapter methods) is a serious infrastructure feat.
Why is it worth noting?
15↑ — public evidence of frontier-scale training running on non-NVIDIA, Chinese hardware: directly relevant to the export-control storyline.