Models Qwen 34b

Tesslate/UIGEN-FX-4B-Preview

Warm
Public
4B
BF16
40960
License: apache-2.0
Hugging Face
Overview

UIGEN-FX-4B-Preview: Frontend Engineer-Tuned Web Generation

UIGEN-FX-4B-Preview is a 4 billion parameter model from Tesslate's UIGEN series, built upon the Qwen3-4B-Instruct-2507 base. This model is uniquely tuned to emulate a frontend engineer, focusing on generating high-quality, web-ready user interfaces across 22 frameworks. Its design prioritizes better visual polish, functional structure, and clean web-ready markup to streamline website development from a single prompt.

Key Capabilities & Features

  • Frontend Engineering Focus: Tuned to behave like a frontend engineer, emphasizing layout rhythm, spacing, and component composition.
  • Web-Only Bias: Specifically trained on curated HTML/CSS/Tailwind snippets and component libraries to produce web-centric outputs.
  • Mobile-First Output: Generates responsive designs with a mobile-first approach.
  • Minimal JavaScript: Defaults to minimal JavaScript, allowing for post-editing for complex interactivity.
  • Efficient Size: At 4 billion parameters, it's small enough for local deployment and fast iteration while maintaining strong structural and visual consistency.
  • High Context Length: Features an effective context length of approximately 64k tokens, suitable for generating practical single-file web pages.

Ideal Use Cases

  • Primary: Generating complete single-file landing pages or web components using frameworks like React, Tailwind, and Javascript, or for Python Frontend WebUI.
  • Secondary: Creating individual component blocks (e.g., hero sections, pricing tables, FAQ accordions) for manual composition into larger projects.
  • Rapid Prototyping: Quickly generating visually polished and structurally sound web interfaces for design exploration and iteration.

Training & Optimization

The model underwent Supervised Fine-Tuning (SFT) with format constraints, followed by instruction tuning and preference optimization focused on style and structure. It was trained on synthetic page specifications and layout constraints to improve visual polish compared to earlier UIGEN releases. Tesslate recommends using a lower temperature (0.4–0.6) for stricter, cleaner markup and a repetition penalty (1.08–1.15) to reduce repetitive classes or markup.