Rumor: DeepSeek V5 in pretraining, ~2T-param sparse MoE trained entirely on Huawei Ascend, no NVIDIA

DeryaTR_ · x · 2026-11-16

Taylor Ogan reports that several friends at Huawei in Shenzhen have quietly moved to DeepSeek in Hangzhou, and claims V5 is already in pretraining: a 2T-parameter ultra-sparse MoE trained on Huawei Ascend 950/950DT with no NVIDIA chips. If it converges, it would be the first frontier model trained entirely on Chinese silicon (unverified).

Reposter DeryaTR argues restricting Jensen Huang/NVIDIA will not slow China's AI catch-up with top US models — if anything, the opposite.

Original post →

More from Infra

Infra channel →