Skip to content
#

zen5

Here are 8 public repositories matching this topic...

用消费级显卡在本地跑动 MoE 架构 30B~35B 大模型:显存与内存混合承担(非专家权重上显存、占大头的专家权重留内存),速度远超纯 CPU;显存不够或没有独显时纯 CPU 兜底。免装 CUDA/Python/任何运行库,模型可替换。

  • Updated Oct 16, 2026
  • C++

Add this topic to your repo

To associate your repository with the zen5 topic, visit your repo's landing page and select "manage topics."

Learn more