OUI-1: World’s First Model for Generative UI

• URL: https://www.threads.com/@vincent_aidev/post/DdRIJovmygi
• Date Saved: 2026-09-15
• Source: Threads (@vincent_aidev)
• Tags: ai-engineering, ai-tools
• Repo: https://github.com/thesysdev/openui (⭐9.3k)
• Blog: https://www.openui.com/blog/oui-1

Summary

Training journey of OUI-1, a DiffusionGemma fine-tune for generative UI, detailing the progressive improvement methodology:

• Base model: 13.0% benchmark — fast (1.6s) but unreliable
• SFT with 700 examples: Score rose to 28.8%, but created a seesaw problem — fixing schema errors worsened wiring errors and vice versa. Speed degraded to 4.3s (output tokens doubled denoising steps)
• RL breakthrough — rejection sampling with parser as reward:
  • Model generates UI code, parser identifies exact errors
  • Near-miss outputs: only fix listed defects, full rewrites rejected
  • Survivors become next round’s training set
  • Simplest form of RL: rejection sampling with parser as reward signal
• Final result: 71.7% benchmark (5.5x base model), speed recovered to 1.9s
  • Beats Gemma 4 31B (46.7%)
  • Second only to Qwen3.8 27B dense
• Generalization: Recipe extended to 27 component libraries; scored 55/60 on unseen AppLess test

Key insight: using a deterministic parser as the reward signal (rather than human feedback or model-based judging) solved the seesaw problem and enabled both error types to decrease simultaneously.