Figure 3 · Muse Spark Open · mixture-of-experts block

Experts work at half width

The expert block of one layer, read top to bottom. 14 of the 256 experts are drawn; which 8 are chosen is illustrative.

The code calls the squeeze and expand "pooler" and "unpooler". The router reads the token at full width, before the squeeze.

latent projections shared by all experts experts chosen for this token experts not chosen a standard step