gWorld-32B

gWorld-32B is Trillion Labs' 33-billion-parameter vision-language world model for mobile GUIs, built on Qwen3-VL-32B and presented in a paper accepted to ICML 2026. Rather than predicting the next screen as pixels, it predicts the next GUI state as executable HTML/CSS code after a user action, which sidesteps the hallucinated and illegible text common in pixel-generation world models.

The model takes a screenshot plus a JSON action (tap, type, and similar actions on a normalized 1000-scale coordinate space) and outputs next-state reasoning followed by renderable HTML, at roughly 0.3 seconds of render time. Trillion Labs reports it beats frontier models up to 12.6x larger on GUI-specific benchmarks, with a 27 percent instruction-accuracy gain over the base Qwen3-VL model.

It is released under Apache 2.0 and is aimed squarely at mobile GUI agent research: AndroidWorld and KApps zero-shot evaluation are the headline benchmarks. A smaller gWorld-8B sibling ships the same design at 9B parameters.