рдкрд░рд┐рдЪрдп
Apache-2.0 open-weight DiffusionGemma fine-tune built specifically to generate interfaces in openui-lang from a supplied component catalog. OUI-1 has 26B total and 4B active parameters, includes local vLLM serving instructions, and is evaluated with parser-verified structural and wiring checks across multiple component libraries.
рд╕рд╛рд░рд╛рдВрд╢
Why OUI-1 Adds New Signal:
1. A model-layer contribution: Ships downloadable weights specialized for UI generation instead of another renderer or protocol
2. Catalog-conditioned output: Reads component signatures from the system prompt and emits validated openui-lang programs
3. Parser-guided training: Combines supervised fine-tuning with rejection-sampled self-distillation and narrowly repaired near-misses
4. Concrete evaluation: Completes 132 of 184 benchmark runs versus 24 for the base model and produces 55 of 60 valid AppLess outputs
5. Local deployment path: Four active billion parameters and documented vLLM serving make on-device or consumer-GPU GenUI a practical research direction
1. A model-layer contribution: Ships downloadable weights specialized for UI generation instead of another renderer or protocol
2. Catalog-conditioned output: Reads component signatures from the system prompt and emits validated openui-lang programs
3. Parser-guided training: Combines supervised fine-tuning with rejection-sampled self-distillation and narrowly repaired near-misses
4. Concrete evaluation: Completes 132 of 184 benchmark runs versus 24 for the base model and produces 55 of 60 valid AppLess outputs
5. Local deployment path: Four active billion parameters and documented vLLM serving make on-device or consumer-GPU GenUI a practical research direction
Tags
openuigenerative-ui-modeldiffusion-language-modelopen-weightsbenchmark