Pinned
AGI achieved locally.
I got Kimi K3’s 2.8T-param MoE (104B active/token) to run on my Mac Studio M3 Ultra with 512GB unified memory.
Mixed Q1/Q4/Q8 quant shrank from 1.56TB to 389.4GiB.
Metal hits 3.36 tok/s.
The future is already here, it's just not very evenly distributed.


