AI BriefingDeepSeek V5 planned for pre-training on Huawei Ascend chips
·2026.11.03 16:07
Key point
DeepSeek V5 is planned to be pre-trained on Huawei Ascend chips, marking the first time a top-tier Chinese model uses domestic hardware for this phase.
Details
- DeepSeek V5 is planned to be pre-trained on Huawei Ascend chips, a shift from previous use of Ascend primarily for inference tasks.
- This move would answer long-standing skepticism about whether China's top-tier large models can be pre-trained on domestic computing hardware.
- DeepSeek has already open-sourced infrastructure components for the Ascend platform, including TileLang, DeepGEMM, FlashMLA, TileKernel, and DeepEP.
- A recent project team completed full-parameter post-training of DeepSeek-V4-Pro (1.6 trillion parameters) on an Ascend 910C cluster, though this excluded the initial pre-training phase.
- DeepSeek anticipates a significant price reduction for the V4-Pro model once the Ascend 950 super-node series enters mass-market availability in the second half of the year.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.
Related
DeepSeek V4.1 Flash Leads Agentic Coding Benchmarks, Narrowing US-China AI Gap to 3%DeepSeek Open-Sources DeepInfer Framework for V4.1 Flash Inference