OUI-1：首个生成式UI开源模型发布

OUI-1是一款基于DiffusionGemma微调的开源模型，专门用于生成用户界面代码1。该模型包含26B个参数，能够在消费级GPU（如RTX 5090）上运行1。在生成式UI基准测试中，OUI-1的得分达到71.7%，相比基础DiffusionGemma的13.0%实现了显著提升，同时超越了Gemma 4 31B（46.7%）和Gemma 4 26B-A4B（29.9%）的表现1。

该模型通过监督微调和自蒸馏两阶段训练实现了可靠性与速度的平衡1。在实际测试中，OUI-1的生成速度从4.3秒降至1.9秒（每输出），输出token数同时增加28%1。在AppLess库上对60个未见过的样本进行测试，OUI-1产生了55个有效输出，而基础DiffusionGemma仅生成了23个1。模型权重已发布在Hugging Face上，遵循Gemma使用条款1。

OUI-1, an open-source model fine-tuned from DiffusionGemma, has been released as the first generative UI system capable of reliably running on consumer-grade GPUs, such as the RTX 5090 1. The 26-billion-parameter model achieved a score of 71.7% on generative UI benchmarks, a dramatic improvement over the base DiffusionGemma's 13.0% 1. This performance gain surpasses competing models, including Gemma 4 31B at 46.7% and Gemma 4 26B-A4B at 29.9% 1.

The model was developed through a two-stage training process combining supervised fine-tuning and self-distillation, balancing reliability with inference speed 1. When tested on 60 previously unseen samples from the AppLess library, OUI-1 produced 55 valid outputs compared to only 23 from the base DiffusionGemma 1. Generation speed improved from 4.3 seconds to 1.9 seconds per output while simultaneously increasing output tokens by 28% 1. Model weights have been made available on Hugging Face under the Gemma terms of use 1.