DeepSeek-V4-Pro Post-Training Completed on Huawei Ascend 910C Chips
The project demonstrates that Chinese-made chips can handle full-parameter post-training of a large-scale model, a step beyond inference that has been a weak point for domestic semiconductors.
Key Facts
- The team used a computing cluster of at least 1,000 Huawei Ascend 910C chips, which are built on SMIC's second-generation 7nm process with Huawei's Da Vinci architecture.
- Huawei claims the Ascend 910C performs comparably to NVIDIA's H100, and the project was completed without using any NVIDIA hardware.
- Post-training is the process that teaches a model to follow instructions and safety rules after its initial pre-training phase.
- The Shenzhen city government described the upgrade as adding complex interchanges and loops to what was previously a one-way road for computing power.
Reporting from 1 source: GIGAZINE.
A Chinese research team including Huawei completed post-training for the DeepSeek-V4-Pro model using at least 1,000 Huawei Ascend 910C chips. The 1.6-trillion-parameter model was trained without NVIDIA hardware, marking progress in China's AI self-reliance amid US sanctions.
The joint team, which includes Huawei, the Shenzhen Loop Area Research Institute, Harbin Institute of Technology Shenzhen Campus, and the Shenzhen Big Data Research Institute, used a computing cluster of at least 1,000 Ascend 910C chips to run the full-parameter post-training. The Ascend 910C is built on SMIC's second-generation 7nm process with Huawei's Da Vinci architecture, and Huawei claims it performs comparably to NVIDIA's H100.
Post-training teaches a model to follow instructions and safety rules after pre-training. The Shenzhen city government described the upgrade as adding complex interchanges and loops to what was previously a one-way road for computing power.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.