Rumor: DeepSeek's rumored single-GPU model may have been trained on Ascend

teortaxesTex · x · 2026-09-30

X user teortaxesTex argues DeepSeek would not open-source kernels hitting 99.8% of hardware limits on GEMM and 98% on MegaMoE without training anything usable end to end. He speculates that at least the rumored small single-GPU model may have been born on Huawei Ascend hardware. Unconfirmed speculation.

Related event: DeepSeek's New Model May Be Trained on Huawei Ascend Chips(3 posts)→

Original post →

More from Infra

Infra channel →