Chinese artificial intelligence company DeepSeek has postponed the release of its next-generation R2 model after encountering persistent problems training it on Huawei's Ascend processors, according to the Financial Times. The delay underscores the technological gap between Chinese and U.S. chips, despite Beijing's push to reduce reliance on Nvidia hardware.
Three people familiar with the matter told the Financial Times that DeepSeek was encouraged by authorities to train R2 on Huawei chips following its R1 debut in January. However, technical failures forced the company to revert to Nvidia hardware for training while attempting to adapt Ascend for inference tasks.
The model's rollout, originally slated for May, has since been postponed. iThome reported DeepSeek has denied renewed speculation that R2 would launch between August 15 and August 30, citing a person familiar with the company's plans. Earlier projections of a March release were also refuted. Reuters sources previously said DeepSeek had accelerated R2's development to improve code generation and multilingual reasoning.
Two people said Huawei dispatched engineers to assist DeepSeek on-site, but the company was unable to complete a successful training run on Ascend. The startup is continuing to work with Huawei to ensure the model functions on Ascend chips for inference.
DeepSeek and Huawei did not respond to requests for comment.
