NVIDIA and AMD Not Required: DeepSeek V4 Trained on Huawei Chips

DeepSeek V4 is using Huawei chips to perform artificial intelligence tasks. According to the company's latest research, Huawei Ascend processors were used for the post-training stage of the V4 Pro model. This is another significant achievement for China's semiconductor industry, as local manufacturers had succeeded in data processing but faced challenges in complex tasks such as model training. Reported by Ixbt.com report .

A group of researchers used a computing cluster consisting of approximately 1,000 Huawei Ascend 910C chips to train the DeepSeek V4 Pro model. Several prestigious institutes in Shenzhen collaborated with Huawei to implement the project. This was a full-parameter training process where the model's capabilities were completely updated without changing its architecture.

In the field of artificial intelligence, inference and post-training are two distinct segments. The first involves running a ready-made model to respond to user queries, while the second involves teaching the model to correctly execute human commands. The post-training stage ensures the model adheres to safety rules and operates based on precise instructions. These steps help China's AI industry achieve independence.

Previously, DeepSeek models were trained on NVIDIA or AMD chips. For example, the DeepSeek V3 model was prepared on a cluster of 2,048 NVIDIA H800 chips. Currently, the DeepSeek startup is preparing to attract investments amounting to 50 billion yuan (approximately $7.4 billion). Competing with models like ChatGPT, DeepSeek demonstrates high efficiency in text generation, coding, and solving logical problems.