Lisan al Gaib· @scaling01 · X·· 3 小时前AI 评分23
AI 导读
一个多月前我把这张图分享给了一位朋友，直到现在我都以为 Meta 可能只是照搬了 Kimi-K2 的架构——1T@32B、61 层 结果 Muse Spark 是 816B@42B、62 层 差不多嘛 :)
正文
I shared this picture with a friend over a month ago and was thinking until now that Meta might have just copied Kimi-K2 architecture at 1T@32B and 61 layers
turns out Muse Spark is 816B@42B and 62 layers
close enough :)
来源：Lisan al Gaib · x.com