| Joint team reports first full-parameter post-training of DeepSeek-V4-Pro on a domestic Ascend cluster A team drawn from Shenzhen Hetao Academy AI Training Platform, Harbin Institute of Technology (Shenzhen), Shenzhen Big Data Research Institute, Huawei GTS, the Computing Product Line and the Huawei 2012 Lab said it had completed the first publicly known third-party full-parameter post-training of DeepSeek-V4-Pro on a Huawei Ascend 910C cluster, Leiphone reported May 29. The project ran more than 1,500 stable training steps over roughly one month, with reported training MFU above 30% and a 14% efficiency gain on a key training operator. The work moves beyond inferenceInferenceRunning a trained AI model to answer a query, as distinct from training it. Inference is where most ongoing compute and energy demand now sits, and it is what chip export rules increasingly target alongside training. and lightweight fine-tuning to full-parameter post-training of trillion-class models on domestic infrastructure. Independent reproduction of the MFU and operator efficiency figures has not been published. |