用 Jev 当教师模型训练 28MB 三值 MoE 玩俄罗斯方块

Shahir @TslShahir原推发布于 2026.09.22

在 X 上打开原推

Given that models playing Tetris is getting popular, we trained our tiny ternary MOE model to play Tetris - Jev style. It does a pretty good job at that. And on the ANE on my M3 Mac, it runs ultra fast. Oh! and it is only 28 MB. Does anyone have a floppy disk? I could make a copy for you :D We used @typesafeai Jev to be the teacher model here, and I think training super tiny decision models with Jev as a teacher is going to be a very popular approach.