DiffusionGemma served locally via SGLang recognizes terrible cat doodles

ChengleiSi · x · 2026-09-24

A user served DiffusionGemma ("Jev") locally with SGLang and found it recognized a terribly drawn cat doodle. It approaches Jev's accuracy with lower latency in local setups and works out of the box, showing diffusion LMs' practical local multimodal understanding.

Original post →

More from Multimodal

Multimodal channel →