Huawei posts trains DeepSeek V4-Pro with 1,000 Ascend 910C chips

Huawei-led team post-trained DeepSeek's 1.6-trillion-parameter V4-Pro model using 1,000 Ascend 910C chips. The training involved full-parameter post-training where all model weights were updated. Work was done by Huawei, Shenzhen Loop Area Institute, Harbin Institute of Technology Shenzhen campus, and Shenzhen Research Institute of Big Data. The Ascend 910C achieved 60% of H100 inference performance in prior tests. Post-training refined the model through instruction-following, safety alignment, and task-specific data.