On October 20th, news emerged indicating that the partnership between DeepSeek and Huawei is evolving from mere software adaptation to a more intricate chip design phase. Insiders with knowledge of the situation reveal that Huawei engineers have been stationed at DeepSeek's facilities for a considerable duration, ensuring that the practical requirements of large-scale model training are seamlessly incorporated into the development of next-generation chips. This integration encompasses crucial elements such as computational task distribution and inter-processor data transmission.

The upcoming focal point of their collaboration is large-scale pre-training. DeepSeek is set to implement Huawei's Ascend 950DT chip for pre-training purposes as soon as the end of this year. However, it's important to note that relevant tests are still in progress, and the precise timeline may undergo adjustments accordingly. Prior to this, DeepSeek has already utilized Huawei's chips to execute model operations and conduct post-training procedures, laying a solid foundation for their ongoing and future endeavors.