返回首页
AI on Medium··行业媒体

Kimi K3 Ran on 8 GB of RAM Without a GPU. The Catch Is 1.7 TB of Storage

中文摘要

Kimi K3 通过 C 引擎从 1.7TB NVMe 存储流式传输大规模 MoE 模型,实现在仅 8GB 内存且无 GPU 的情况下运行,但速度较慢。

English Summary

Kimi K3 runs a massive MoE model on 8GB RAM without a GPU by streaming 1.7TB from NVMe via a C engine, though performance is slow.

原文节选

How a tiny C engine streams a frontier-scale MoE model from NVMe, why the minimum-memory mode remains slow and what the experiment… Continue reading on Medium »