nbeerbower/Gemma4-Glamour-31B

VISIONConcurrent Unit Cost:2Model Size:31BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 20, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

nbeerbower/Gemma4-Glamour-31B is a 31-billion parameter language model fine-tuned from google/gemma-4-31B-it, specifically designed for generating web-design code. This model excels at transforming natural language briefs into self-contained front-end implementations, outputting HTML, CSS, and JavaScript. It is part of the Glamour project by Schneewolf Labs, focusing on web-design generation and critique. The model's primary strength lies in its ability to produce functional web interfaces from textual descriptions.

Loading preview...

Model Overview

nbeerbower/Gemma4-Glamour-31B is a specialized 31-billion parameter language model, fine-tuned from Google's gemma-4-31B-it architecture. It is developed as part of the Glamour project by Schneewolf Labs, which focuses on web-design generation and evaluation. The model's core function is to generate complete front-end web designs (HTML, CSS, JavaScript) based on natural language prompts.

Key Capabilities

  • Web-Design Generation: Translates natural language descriptions into functional web-design code.
  • Structured Output: Provides self-contained front-end code in three distinct fenced blocks: html, css, and js.
  • Fine-tuned Performance: Trained using QLoRA SFT on approximately 445 brief-to-implementation pairs, specifically from the Opus-authored subset, demonstrating improved performance even on designs originally authored by other models.

Training Details

The model underwent QLoRA SFT (r=64, α=32) and was merged to full weights. It was trained on a dataset of web-design briefs and their corresponding implementations. This release represents a phase-1 generator, captured at an early-stop checkpoint, indicating ongoing development in format adherence.

Usage Considerations

This model is designed for generating web interfaces. While format adherence is still developing, it serves as a strong v1 for web-design automation. It can be integrated with openai compatible chat clients, and a Q4_K_M GGUF version is noted to run efficiently on llama.cpp.

Finetunes
1 models