Any plans for training a dedicated A2UI-native model?

[binIoter]

Hi team 👋,
We've been following the A2UI project with great interest, especially the recent Express proposal, which takes a compelling approach to reducing token consumption and lowering response latency.
This got us thinking about a more fundamental question regarding the long-term roadmap:
The Question
Is there any plan to train a dedicated, A2UI-native model — one that natively understands and generates A2UI structures, rather than relying on general-purpose LLMs guided by prompts?
Our reasoning:
● Current general-purpose LLMs understand A2UI through prompt engineering and in-context instructions. While functional, this approach has inherent ceilings in terms of output consistency, token efficiency, and latency.
● A dedicated A2UI-native model could fundamentally change the equation:
● Natively internalize the A2UI schema and component semantics, eliminating the overhead of in-context instructions
● Dramatically reduce token usage and latency — Express support would simply be a natural, lightweight configuration on top of this
● Improve output fidelity for complex layouts, nested component hierarchies, and edge cases
● Potentially enable smaller, faster, locally deployable model variants purpose-built for UI generation tasks
The way we see it, Express is a format-level configuration. The real leverage comes from having a model that thinks in A2UI from the ground up.
A few follow-up questions:
● Is there any dataset curation or annotation effort currently underway to support this direction?
● Are there plans for community involvement — e.g., contributing UI-to-A2UI training pairs or evaluation benchmarks?
● Is this something being actively explored, or is the current strategy to remain model-agnostic and rely on prompt-based approaches?
Even a high-level answer would be really helpful for understanding where the project is headed. Thanks for building this! 🙌