Introducing OUI-1: world's first model for Generative UI

Summary

OUI-1 is a 26B-A4B (4B active parameters) DiffusionGemma finetune that generates user interfaces in OpenUI Lang, a streaming-first UI language using up to 67% fewer tokens than JSON. It scores 71.7% on the Generative UI Benchmark versus 13.0% for the base model, runs on consumer GPUs (RTX 5090, FP8), and is released open-weight on Hugging Face under Apache 2.0.

Keywords

generative UI · diffusion language models · self-distillation · open weights · OpenUI Lang

Topics

generative UI, diffusion language models, self-distillation, open weights, OpenUI Lang

Research notes

• Method: Two-stage training: supervised fine-tuning on OpenUI Lang examples, then rejection-sampled self-distillation where the OpenUI Lang parser acts as a verifiable reward — the model generates programs, the parser keeps valid ones, near-misses are repaired, and survivors become the next training set. The recipe was applied across 27 component libraries.
• Key findings: 71.7% on Generative UI Benchmark (base DiffusionGemma 13.0%); Beats Gemma 4 31B (46.7%); only dense Qwen3.8 27B (78.8%) scored higher; Self-distillation cut generation time from 4.3s to 1.9s per output while reducing both schema and wiring errors simultaneously; Generalized to unseen AppLess component library: 55/60 valid outputs vs 23/60 for base model
• Limitations: Still short of the sub-second latency target (~1.9s/output); the hypothesis for why self-distillation works (gradient concentration on repaired statements) is stated as not isolated/verified.
• Blog post date not shown on page; published on or before 2026-09-08 (Discord discovery date). Model weights linked on Hugging Face.