WEBGEN-4B-Preview: A Specialized Web Generator
Tesslate's WEBGEN-4B-Preview is a 4 billion parameter model specifically engineered for generating web content. Built upon Qwen/Qwen3-4B-Instruct, this model is fine-tuned to produce production-lean websites, prioritizing semantic HTML, consistent spacing, and modern component blocks like heroes, grids, pricing sections, and FAQs.
Key Capabilities
- Web-Only Generation: Focuses exclusively on creating complete, single-file websites (landing pages, marketing pages, simple documentation) and reusable component blocks.
- Output Format: Generates clean HTML/CSS, with a strong preference for Tailwind CSS for styling, ensuring responsive and modern designs.
- Efficiency: At 4 billion parameters, it's designed to be small enough for local execution and fast iteration, making it suitable for developers seeking quick prototyping.
- Structured Output: Emphasizes strong structural consistency and opinionated layouts, preferring mobile-first output and avoiding external JavaScript by default.
- Training Focus: Trained on a mixture of curated HTML/CSS/Tailwind snippets, component libraries, and synthetic page specifications, with an objective to reward semantic structure, spacing rhythm, and responsiveness.
When to Use This Model
- Rapid Prototyping: Ideal for quickly generating website layouts or specific UI components from text prompts.
- Semantic HTML Needs: When the primary requirement is well-structured, semantic HTML code.
- Tailwind CSS Projects: Particularly useful for projects that leverage Tailwind CSS for styling.
- Resource-Constrained Environments: Its 4B size makes it suitable for deployment on laptops or environments with limited GPU memory, with BF16 requiring 8.05 GB VRAM and GGUF Q4_K_M needing only 2.5 GB.
For optimal results, users are advised to use a lower temperature (e.g., 0.4–0.6) for stricter, cleaner markup, and a max_new_tokens between 1200–2500, as single-file sites often fit within 1500 tokens.