Chinese scientists are moving from inference to training AI models on Huawei accelerators

Huawei’s Chinese chips successfully completed post‑training of the DeepSeek‑V4‑Pro model

*According to a South China Morning Post report, China managed to use Huawei Ascend 910C chips for full fine‑tuning of the large AI model DeepSeek‑V4‑Pro. This is an important step in developing the domestic semiconductor industry, which aims to move from simple inference to complex training amid tightening U.S. sanctions.*

What happened
Project launch | Research team from Huawei Technologies and partners (Shenzhen Ring Road Institute, Shenzhen Campus of Harbin Institute of Technology, Shenzhen Big Data Research Institute) started with the goal of training the DeepSeek‑V4‑Pro model. Scale | The model contains 1.6 trillion parameters and was launched on a cluster of at least 1,000 Ascend 910C chips. Post‑training | Fully “parameterized” fine‑tuning: the entire architecture of the model was updated without compromises, not just individual layers.

Why it matters
* Until recently Chinese chips were successfully used for inference – the process where the model simply answers queries.

* Now, thanks to the new project, the model can self‑reflect and adjust during operation, adding “complex overpasses and loops” to the previously one‑way data flow. This increases computational and communication demands but makes the system more autonomous.

Expected outcomes
* The Shenzhen government stated that the research will help increase the self‑sufficiency of China’s AI industry and strengthen the position of domestic chips amid external restrictions.