OpenUI's OUI-1 model writes interfaces instead of JSON

OpenUI fine-tuned a 26B-parameter model, 4B active, to write user interfaces directly instead of describing them in JSON. The model, called OUI-1, is built on DiffusionGemma. It outputs a custom format called openui-lang that streams progressively as it generates, so a UI can start rendering before the full response closes.

That matters for anyone paying per token or streaming UI to a slow connection, since OpenUI says the format cuts token usage by up to 67% against JSON output. OUI-1 scores 71.7% on the Generative UI Benchmark, against 13% for the base DiffusionGemma model it started from.

The harder test is generalization. OpenUI ran 60 independent requests against a component library the model had never seen during training. OUI-1 returned 55 valid outputs. The base model returned 23. That's the gap that would show up in production, where a model built for known component libraries meets ones it wasn't trained on.

Speed and access back up the pitch. OpenUI says the model runs on a single RTX 5090 in 1.9 seconds, and the weights are free on Hugging Face under the Gemma terms, so anyone can pull it down and test the claims directly rather than take a benchmark screenshot on faith, per OpenUI's OUI-1 writeup.

What the post doesn't cover is cost to reproduce, how openui-lang handles UI patterns outside the benchmark's scope, or how it holds up against larger general-purpose models asked to do the same job. Free weights make that comparison easy to run, and somebody will.