A2UI separates model inference syntax from the wire protocol with Express and controlled decisions
On October 27, the Google A2UI team introduced swappable inference formats: a model can emit compact A2UI Express and let the SDK deterministically compile it against a component catalog into standard A2UI messages, while fixed-shape screens can use decision models to select predefined layouts.
Why it matters
Generative-UI latency and validity depend on what the model is asked to emit as well as the model itself. Separating a model-friendly intermediate representation from the stable wire protocol can reduce tokens, swap inference engines and retain host catalog validation without teaching every renderer a new model syntax.
A closer look
A2UI Express describes interfaces through line-oriented declarations, positional arguments, data bindings and events. Implementations are available in Python agent SDK 0.9.0, TypeScript and Dart. The SDK derives prompts and a compiler from catalog schemas, validates the result, and emits standard A2UI. Google reports up to roughly three times fewer tokens and up to three times lower latency than direct JSON, based on a 51-task comparison on Gemini 3.5 Flash. A controlled path asks Jev about roughly thirty layout decisions, assembles the component tree in code and uses a small model only for text. In the 11-prompt Apparite reference run, first components appeared in about 92–298 ms, pages completed in 0.8–1.7 s, and all 11 outputs validated as A2UI.
Who this is for
Teams designing generative-UI protocols, component catalogs, model output formats, low-latency screen generation and on-device inference
Original sources
- Google Developers Blog, A2UI specification and reference implementation ↗developers.googleblog.com
- A2UI Express technical specification ↗a2ui.org
- A2UI Python agent SDK 0.9.0 changelog ↗github.com
- Apparite reference results and known limits ↗github.com
Revisions & corrections (1)
— Official A2UI announcement, specification, SDK changelog and reference implementation jointly establish a new inference-format layer, measured tradeoffs and explicit limitations.