Deepseek Delays AI Model Due to Huawei Chip Issues

Chinese AI company Deepseek has reportedly postponed the release of its latest AI model after encountering difficulties with Huawei's Ascend chips.

According to the Financial Times, Chinese regulators advised Deepseek to transition from Nvidia's leading chips to Huawei's Ascend processors following the release of its R1 model in January. However, the plan faced challenges as Deepseek experienced ongoing technical issues with the Ascend chips during the training of its R2 model. Even with Huawei engineers present, the team was unable to complete a successful training run.

These complications forced Deepseek to revert to Nvidia chips for the training process, delaying the model's launch from May and allowing competitors to gain an advantage. As a temporary solution, Deepseek now utilizes Nvidia hardware for training while employing Huawei's Ascend chips for less intensive inference tasks. Industry sources indicate that Chinese chips still lag behind Nvidia in terms of stability, connectivity, and software quality.

Despite these setbacks, Deepseek has released an updated version of its V3 model. The Register reports that the new V3.1 was trained using a special data type called UE8M0 FP8. In a WeChat post, Deepseek stated that this data type is designed for a new generation of domestically produced chips, which are expected to be released soon.

This development suggests that more advanced Chinese accelerators may be forthcoming. Huawei's current top chip, the Ascend 910C, does not natively support the FP8 data type. The shift from the previously used E4M3 format appears to focus more on future hardware compatibility than efficiency. V3.1 builds on an earlier V3 checkpoint, adding a hybrid reasoning mode.