Have you ever tried bigger model on this methods?

[447162527]

Hi neovateai, this is a excellent work on UI-related tasks. And I found that the UI-UG of 7B parameters outperforms many models on referring and grounding tasks. But it seems still improvement on generation tasks. I want to ask that have you ever tried bigger size of model on this methods (Qwen-2.5-vl-14b, 32b, internVL)?

Best wishes.

[AirBobby]

Thank you very much for your recognition of our work. Previously, considering the real-time performance requirements of generative inference, we only experimented with 7B and 3B models, and the 7B model was released.
But improving generation quality will remain a key focus in our next phase, including constructing higher-quality UI generation training data (which currently may appear to be a weak point). Additionally, we plan to explore larger MoE models in the future, such as the latest Qwen3-VL-30B (A3B).

[seasonmac]

Is it possible to leverage community power to build datasets?