The Information reported that Huawei engineers are stationed at DeepSeek offices and are incorporating model-training requirements into next-generation Ascend chip design, according to a BlockBeats flash.
From failed training to co-design
DeepSeek previously failed to train its R2 model on Huawei Ascend chips, with the Financial Times attributing the problems to chip stability, cross-card communication and software support. DeepSeek returned to Nvidia chips to complete that training, the report said.
Domestic-chip roadmap
BlockBeats said DeepSeek open-sourced an Ascend version of TileLang and related low-level tools in late September. Huawei reportedly plans to prioritize most Q1 2027-ready Ascend 960DT supply for DeepSeek, which plans to start pretraining with 950DT by year-end at the earliest. Tests have used more than 1,000 chips, while full pretraining may require tens of thousands; some cluster software remains unfinished and DeepSeek is still seeking Nvidia chips.
Source: BlockBeats flash, citing The Information and the Financial Times.