⚡ STARK-WEB-12B-GGUF

Local GGUF Quantizations for the Premium AI UI/UX Designer & Frontend Model (v1.0.0)

---

🔍 Model Overview

This repository contains local GGUF quantizations of STARK-WEB-12B, a specialized fine-tune of Google's Gemma 4 12B architecture designed to act as a premium UI/UX designer and frontend engineer. It generates complete, visually stunning, single-file web applications combining semantic HTML5, modern CSS, and interactive JavaScript.

These files are fully compatible with local inference engines like LM Studio, Ollama, llama.cpp, Llamafile, and Faraday.

---

⚠️ Quantization Status & Known Issues

⚡ GGUF File Summary & Formatting Notes

File Name	Quantization	Status & Recommendations
...-f16.gguf	F16 (16-bit)	Perfect. Maximum reasoning quality. Recommended if memory allows (~24 GB VRAM/RAM).
...-q8_0.gguf	Q8_0 (8-bit)	Highly Recommended. Near-zero degradation in CoT structuring and UI aesthetics. (~13 GB).
...-q4_k_m.gguf	Q4_K_M (4-bit)	⚠️ Known Formatting Issue. The q4_k_m version occasionally introduces syntax/formatting mistakes that can break page rendering. We are actively working to fix this. If you experience layout breakage, please use the Q8_0 or F16 files.

---

🛠️ How to Run Locally

Prompt Template & Inference Settings

The model was fine-tuned without a system prompt. Use the standard Gemma format (<|turn|>). The chat template automatically injects the <|think|> token on system level.

Expected Template Structure:

<|turn>system
<|think|>
<turn|>
<|turn>user
Make a responsive, beautiful game of Tetris in 1 file
<turn|>
<|turn>model
<|channel>thought
1. Understand the Goal:
...

💡 Recommendation for Broken Output (Second Pass)

If you encounter layout bugs, copy the output code and feed it back to the model (or a larger model like Gemini 3.1 Flash lite/Claude etc,) with this correction prompt:

"Review the generated code above. Identify any rendering or logical errors, and output a corrected, fully functional, and complete single-file version."

---

📜 Citation & License

This model is licensed under Apache 2.0.

@misc{situus2026starkweb,
  author = {Situus},
  title = {STARK-WEB-12B-GGUF: Local GGUF Quantizations for STARK-WEB-12B},
  year = {2026},
  publisher = {Hugging Face},
  journal = {Hugging Face Model Hub}
}