Z.ai Paper Points to GLM-Omni, a Model That Draws and Reads
A Zhipu AI and Tsinghua paper reports text QA rising from 12.57% to 26.86% on GLM-Omni.
Z.ai’s next model may generate images as well as read them.
“ZAI is working on an OMNI model,” @MiaAI_lab posted early Monday. “It's called GLM-Omni in their new paper, a model that can both understand and generate images. This could be huge for local AI.”
The post she quoted comes from @volatilemarkts. It names the paper: “Can Visual Generation Expand What Unified Multimodal Models Know?” Its authors are Wenkai Li, Wenyi Hong, Zhuoyi Yang and Jie Tang of Zhipu AI and Tsinghua University. It is on arXiv as 2610.21459.
The researchers built a benchmarking protocol called Knowledge Transfer Probing (KTP), @volatilemarkts writes. They taught models new factual concepts strictly through image generation tasks.
At first the models showed “modal aphasia.” A model could draw a complex concept in pixels and still struggle to state it in text.
| Finding | Result |
|---|---|
| Interleaved image-text alignment | Text QA on GLM-Omni rises from 12.57% to 26.86% |
| Implicit cross-entity generalization | Training on a small subset of visual entities lifts recall on unseen entities in the same category |
@volatilemarkts guesses at “GLM 5.4 or 5.5, possibly GLM Omni.” He asks Z.ai to give @MiaAI_lab and others early access for zero day support on Sparks and Mac Studios.
Sources
- @MiaAI_lab on X · Oct 19, 4:58 AM PT