OUI-1: first open-weight model for generative UI, runs on a 5090
A DiffusionGemma fine-tune that writes app screens at 700 tok/s on one consumer GPU, scoring 71.7% on the GenUI bench. Plus: D2's TALA goes open.
Copy markdown
What it is: a 4B-active DiffusionGemma fine-tune
OUI-1 is a 26B-parameter MoE (4B active) that Thesys fine-tuned from Google's DiffusionGemma to write whole app screens. It runs on a single RTX 5090 at FP8 and streams 256-token blocks at ~700 tok/s—about a second per light screen. Weights are on Hugging Face under the Gemma license.
OpenUI Lang: 67% fewer tokens than JSON
Instead of emitting JSON, it writes OpenUI Lang, a declarative UI DSL that streams and cuts token count by up to 67%. It targets 27 component libraries out of the box, so generated screens map onto components you already ship.
The scores: 71.7% on the GenUI bench
OUI-1 hits 71.7% on the Generative UI Benchmark (46 screen briefs, validator-scored)—5.5x its 13% base model and past Gemma 4 31B's 46.7%. Only Qwen3.8 27B beats it (78.8%), and it does so with 27B dense params versus OUI-1's 4B active.
Build this weekend: software that draws its own UI
Point it at a component library and generate dashboards from ad-hoc queries, conversational analytics, or per-user screens on the fly—sub-second on one GPU. Serve it locally with vLLM 0.24+, or hit Thesys's Cloud Gateway endpoint if you'd rather not host.
The catch: Gemma terms, and per-request drift
It's open weights, not OSI-open—the Gemma Terms of Use carry usage restrictions. And because each screen is generated fresh, early testers flag UI that shifts between sessions, which complicates support and muscle memory.
Elsewhere: TALA, D2's layout engine, goes open
Terrastruct open-sourced TALA—the auto-layout engine behind D2, built for software-architecture diagrams—under MPL-2.0. If you write diagrams as code, its good layout is now free.