Huawei-Led Team Post-Trains DeepSeek's 1.6T Parameter Model Using 1,000 Ascend 910C Chips
A Huawei-led team has reportedly post-trained DeepSeek's large language model, which contains 1.6 trillion parameters, using approximately 1,000 Ascend 910C chips. The training was conducted entirely on Huawei's Ascend hardware ecosystem, which has become a key pillar of China's strategy to develop competitive AI systems using domestically produced chips amid semiconductor export restrictions.
The Ascend 910C is Huawei's latest AI accelerator designed to compete with high-end graphics processing units in training large-scale machine learning models. By scaling the post-training process across 1,000 of these chips, the team demonstrated a path toward handling computationally intensive AI workloads without relying on foreign hardware suppliers.
This development underscores the continued competition in the global AI landscape, where both hardware and algorithmic advances are driving rapid changes. China's AI ecosystem has increasingly focused on building independent capabilities, with Huawei's Ascend series playing a central role in that strategy.