OUI-1 Released: DiffusionGemma-based Generative UI Model Achieves 71.7% on Benchmark

Thesys released the OUI-1 model, fine-tuned based on DiffusionGemma. This model achieved a score of 71.7% on the Generative UI Benchmark, improving performance by 5.5x compared to the base model. With a 26B parameter structure (active parameters 4B), it can run in FP8 format on consumer GPUs (such as the RTX 5090).

Training Strategy and Performance Improvements

OUI-1 uses the OpenUI Lang protocol, reducing token usage by up to 67% compared to JSON and supporting streaming generation. Training proceeded in two phases.

• Phase 1 (SFT): LoRA fine-tuning was performed using examples written by large models, but generation time increased and parsing failures occurred. Narrowing the scope to a single library raised the score to 28.8%, but a trade-off existed between Schema errors and Wiring errors.
• Phase 2 (Self-distillation): Self-learning was conducted by applying parser-based validation rewards and Rejection Sampling. This reduced generation time from 4.3 seconds to 1.9 seconds and simultaneously decreased Schema and Wiring errors, recovering the score to 57.1%.

Benchmark Results and Comparison

Finally, after additional training on 27 Component Libraries (Phase 3), a final score of 71.7% was recorded.

• Comparison Models: Qwen3.8 27B (78.8%), Qwen3.6 27B (68.5%), Gemma 4 31B (46.7%), etc.
• Features: Achieved the highest performance among models with 4B active parameters or fewer, showing performance close to Dense 27B-class models.
• Generalization Ability: On the AppLess library not used in training, it generated 55 valid outputs compared to DiffusionGemma (23), proving generalization performance.

Deployment and Future Plans

Model weights were released on Hugging Face under the Gemma Terms of Use, and benchmark code and original generations can be found on GitHub. Development will continue with goals of local execution on personal devices, state management interface generation via OpenUI Lang 0.5, and low-latency local generation within 1 second.