1 · Architectures — common core, different bets · 19/38
MoE — 384 exist, 8 fire
VRAM follows total params. FLOPs follow active.
playtop-kexperts
nvidia-smi says—
the SMs it touches really are busy
active parameters—
only k of N experts fired this token
K2 needs 8×H100 even though ~32B compute per token. VRAM follows total.