UIGEN-T3 — Advanced UI Generation with Hybrid Reasoning
Tesslate’s next-gen UI model, built for thoughtful design.
Demos
Explore New UI generations: 📂 https://uigenoutput.tesslate.com
Join our Discord: https://discord.gg/GNbWAeJ4 Our Website: https://tesslate.com
Quick Information
- UI generation model built on Qwen3 architecture
- Supports both components and full web pages
- Hybrid reasoning system: Use
/thinkor/no_thinkmodes - Powered by UIGenEval, a first-of-its-kind benchmark for UI generation
- Released for research, non-commercial use. If you want to use it commercially, please contact us for a pilot program.
Model Details
- Base Model: Qwen/Qwen3-14B
- Reasoning Style: Hybrid (
/thinkand/no_think) - Tokenizer: Qwen default, with design token headers
- Output: Components + Full pages (with
<html>,<head>) - Images: User-supplied or placehold.co – no images in the dataset due to licensing concerns.
- License: Research only (non-commercial). Contact us for enterprise use cases.
Reasoning System
UIGEN-T3 was trained using a pre/post reasoning model architecture.
You can explicitly control the reasoning mode:
/think→ Enables guided reasoning with layout analysis and heuristics./no_think→ Faster, raw code generation.
Outputs also include design tokens at the top of each generation for easier site-wide customization.
Inference Parameters
Please use 20k context length to get the best results if using reasoning.
| Parameter | Value |
|---|---|
| Temperature | 0.6 |
| Top P | 0.95 |
| Top K | 20 |
| Max Tokens | 40k+ |
Evaluation: UIGenEval Framework
UIGenEval is our internal evaluation suite, designed to bridge the gap between creative output and quality assurance. (Learn more in our upcoming paper: "UIGenEval: Bridging the Evaluation Gap in AI-Driven UI Generation" - August, 2025)
UIGenEval evaluates models across four pillars:
- Technical Quality — Clean HTML, CSS structure, semantic accuracy.
- Prompt Adherence — Feature completeness and fidelity to instructions.
- Interaction Behavior — Dynamic logic hooks and functional interactivity.
- Responsive Design — Multi-viewport performance via Lighthouse, Axe-core, and custom scripts.
This comprehensive framework directly informs our GRPO reward functions for the next release.
Example Prompts to Try
make a google drive clonebuild a figma-style canvas with toolbarcreate a modern pricing page with three plansgenerate a mobile-first recipe sharing app layout
Use Cases
| Use Case | Description |
|---|---|
| Startup MVPs | Quickly scaffold UIs from scratch with clean code. |
| Design-to-Code Transfer | Figma (coming soon) → Code generation. |
| Component Libraries | Build buttons, cards, navbars, and export at scale. |
| Internal Tool Builders | Create admin panels, dashboards, and layout templates. |
| Rapid Client Prototypes | Save time on mockups with production-ready HTML+Tailwind outputs. |
Limitations
- No Bootstrap support (planned).
- Not suited for production use — research-only license.
- Responsive tuning varies across output complexity.
Roadmap
| Milestone | Status |
|---|---|
| Launch Tesslate Designer | 2 days |
| Figma convert | |
| Bootstrap & JS logic | |
| GRPO fine-tuning | |
| 4B draft model release | Now |
Technical Requirements
- GPU: ≥16GB VRAM for 14B inference on GGUF.
- Libraries:
transformers,torch,peft. - Compatible with Hugging Face inference APIs and local generation pipelines.
Community & Contribution
- Join our Discord: https://discord.gg/GNbWAeJ4
- Chat about AI, design, or model training.
- Want to contribute UIs or feedback? Let’s talk!
Citation
@misc{tesslate_UIGEN-T3,
title={UIGEN-T3: Hybrid Reasoning for Robust UI Generation on Qwen3},
author={Tesslate Team},
year={2025},
publisher={Tesslate},
note={Non-commercial Research License},
url={https://huggingface.co/tesslate/UIGEN-T3}
}