On August 14, 2025, the Financial Times reported that DeepSeek encountered persistent technical issues while training the R2 using Ascend chips, prompting a switch to NVIDIA chips for training, and using Huawei chips for inference.
These issues are the main reason for the postponement of the model’s release date from May.Training is the process of optimizing model parameters using massive datasets, while inference refers to the process of predicting or generating outputs from new input data after the model has been trained (such as generating answers in a question-answering system).
Huawei has sent “a team of engineers” to DeepSeek’s office to assist in the development of the R2 model using Ascend chips.
Despite the team being stationed on-site, DeepSeek still cannot successfully train on the Ascend chips.
The reasons include insufficient stability of the Ascend platform during training tasks, inadequate software and hardware support, and slow communication speeds between chips.
DeepSeek is still collaborating with Huawei to ensure that its model is compatible with Ascend to complete inference tasks.
DeepSeek’s founder, Liang Wenfeng, has expressed dissatisfaction with the progress of the R2 and has been striving to invest more time in developing an advanced model that can keep the company at the forefront of the AI field.
The release of the R2 has also been delayed due to the data annotation time for the updated model exceeding expectations.
