wesjos/mario-html-4B

VISIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Oct 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

wesjos/mario-html-4B is a 4.5 billion parameter visual-UI language model fine-tuned from Qwen3.5-4B. It specializes in generating structured HTML components like bento grids, metric cards, and steppers, wrapped in a single root, instead of plain Markdown. This multimodal model supports both text-to-HTML and image-to-HTML generation, and is designed to be reasoning-native without requiring a system prompt. Its primary use case is to provide visually rich, structured outputs for applications requiring UI elements directly from an LLM.

Loading preview...

Mario-HTML-4B: Visual-UI Language Model

Mario-HTML-4B is a 4.5 billion parameter language model, fine-tuned from Qwen3.5-4B, that generates responses in semantic HTML components rather than traditional Markdown. This model is an experiment to explore the utility of HTML as an output format for LLMs, providing structured UI elements directly.

Key Capabilities

  • HTML-Native Output: Generates responses using a predefined set of HTML components (e.g., <div class="ai-view">, <div class="card">, <div class="stepper">, <div class="callout">) for structured, visually rich outputs.
  • Multimodal: Supports both text-to-HTML and image-to-HTML generation, allowing for visual input processing.
  • Reasoning-Native: Designed to handle reasoning tasks effectively without the need for an explicit system prompt.
  • Structured UI Elements: Outputs include bento grids, metric cards, steppers, and callouts, facilitating direct integration into web interfaces.

Good For

  • Generating UI Components: Ideal for applications that require an LLM to directly produce structured, styled web UI elements.
  • Visual Content Description: Useful for describing images and outputting the description in a structured HTML format.
  • Interactive Applications: Can serve as a backend for web applications where the LLM's output is immediately rendered as part of the user interface.
  • Experimentation with LLM Output Formats: Provides a unique approach to LLM interaction by focusing on HTML as the primary output medium.

Finetunes
2 models