mousepotato (@octg): "2.8T 参数的 Kimi K3 有人说不用先砸一张顶级卡也能碰项目叫 colibri 思路很粗暴也很大胆：专家权重不整模塞内存需要哪个 expert 再从固态硬盘流进来 MoE 本来就是「每次只激活一小撮」那磁盘延迟就变成可交易的工程问题而不是绝对禁区能跑 ≠ 很爽但「能不能在消费级硬件上试」这条线又被往下推了一截 https://github.com/JustVugg/colibri"

Make money doing the work you believe in

章鱼花园’s Newsletter

2.8T 参数的 Kimi K3

有人说不用先砸一张顶级卡也能碰

项目叫 colibri

思路很粗暴也很大胆：

专家权重不整模塞内存

需要哪个 expert

再从固态硬盘流进来

MoE 本来就是「每次只激活一小撮」

那磁盘延迟就变成可交易的工程问题

而不是绝对禁区

能跑 ≠ 很爽

但「能不能在消费级硬件上试」

这条线又被往下推了一截

github.com

GitHub - JustVugg/colibri: Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦 - JustVugg/colibri

Aug 14

at

5:54 AM

#nojs-banner { position: fixed; bottom: 0; left: 0; padding: 16px 16px 16px 32px; width: 100%; box-sizing: border-box; background: red; color: white; font-family: -apple-system, "Segoe UI", Roboto, Helvetica, Arial, sans-serif, "Apple Color Emoji", "Segoe UI Emoji", "Segoe UI Symbol"; font-size: 13px; line-height: 13px; } #nojs-banner a { color: inherit; text-decoration: underline; } This site requires JavaScript to run correctly. Please turn on JavaScript or unblock scripts