A China Telecom lab reports Xing 4 training on Ascend chips

China Telecom's Xingchen AGI Lab said in a Nov. 17 preprint that it pretrained a 285-billion-parameter Xing 4 model, with 17 billion parameters active per token, on 8,192 Ascend 910C cards. The lab said the work was done with Huawei and Harbin Institute of Technology, Shenzhen, and that it open-sourced a 29-billion-parameter model trained on Ascend with MindSpore. Its 4.68-times prefill comparison is a simulation against an MLA-only baseline, and the authors say simulation can be more optimistic than full deployment.

/ Sources

/ About this story

Compiled by Venture Atlas from the sources above, using automated AI-assisted research. This is a summary of reporting published elsewhere, not original reporting - follow the source links for the full account. See our editorial standards.

Something wrong here? Email apostructura@gmail.com and we will fix it.

/ Related