Architecture families

Choosing the right one, and why it matters.

The selector at the top left sets the family. It decides which blocks the palette offers and how the analysis treats them. A mixture-of-experts model left on "transformer" has its experts counted once instead of per expert.

• Transformer — encoder, decoder and encoder-decoder models.
• CNN — convolutional vision models.
• MoE — mixture of experts, with routers and expert blocks.
• SSM — state-space models such as Mamba and RWKV.
• Diffusion — denoisers, U-Nets and DiTs.
• GNN — graph networks.

There are 319 blocks across these 6 families, and 80 templates built from them ship with the studio; the Templates panel adds those the NEURAX service offers.

Page 16 of 33. This guide is in English for now. It is the same guide that is built into the NEURAX desktop app (the ? button, or F1). Found a mistake? Tell us.