Qwen3.5-4B-3D-Html

Qwen3.5-4B-3D-Html

A 4.2B text-generation model, fine-tuned from Qwen3.5-4B, that writes a 3D scene as one self-contained HTML file using Three.js. Open the file in a browser and rotate the model with the mouse.

Made by mondk on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: Qwen3.5 4B 3D Html is a small text-generation model fine-tuned from Qwen3.5-4B. It writes a 3D scene as one self-contained HTML file built with Three.js. You describe an object, save the reply as a web page, and open it in a browser. The card says no extra libraries need installing. A set system prompt asks for a white background, no buttons or text, scroll to zoom and the right mouse button to rotate.

Who it suits: It suits web developers, teachers and makers who want a quick 3D preview from a text description, and people who test small models on code-made 3D.

Industry | Who uses it | How they use it | Example
Web development | Front-end developers | Generate a Three.js scene as one file | Ask for a simple object and open the file in a browser
Education | Teachers of 3D graphics | Show how a shape is built from code | Generate a shape and read its Three.js code
Hobby and prototyping | Makers | Sketch 3D ideas quickly | Request a small object and rotate it with the right mouse button

If you need polished, detailed models or exact results, you can skip it. The card says it makes minor errors and that its training data is small.

Where to use it:

• A runner for Safetensors or GGUF models, since the repository offers both
• A web browser, to open the generated HTML file
• A computer with room for about 8.4 GB in BF16, or about 2.1 GB at 4-bit, which is an estimate
• Windows, macOS or Linux with Python, to download the files

Use case | Who | What they get back
3D scene from text | Makers | One HTML file holding a Three.js scene
Quick prototype | Developers | A starting point you can edit
Teaching | Teachers | Readable code for a simple 3D object

How to set it up: What you need first:

• Python and the Hugging Face command line
• A way to run the model, either from the Safetensors files or from the Q6_K GGUF file the card lists
• Room for the weights

1. Download the model files.
2. Load the model in your runner and set the system prompt from the card.
3. Apply the sampling settings, and set a high limit for output tokens.

pip install -U huggingface_hub
hf download mondk/Qwen3.5-4B-3D-Html

How to use it: Set the system prompt below, then describe the object you want in your message. The model replies with one HTML file. Save it with an html ending and open it in a browser. Scroll to zoom and hold the right mouse button to rotate.

You are a text-to-3D model generator. You will write a single HTML file using Three.js to create a 3D model.
The output must be a single, self-contained HTML file (all JavaScript and CSS inline in that one file). Return only the complete HTML code.
Load Three.js and OrbitControls from a CDN (e.g. cdnjs or jsDelivr) using a pinned version.
The background must be white.
Do not add any additional elements (buttons, text, etc.) to the interface; the view should be exclusively for observing the 3D model.
Controls: Use the scroll wheel to zoom in/out, and the right mouse button to rotate the camera.

Title: Sampling settings from the card
Setting | Value
Temperature | 0.65
Top P | 0.95
Top K | 30
Repeat penalty | 1.05

Tool | What it does
Three.js | Draws the 3D scene in the browser
OrbitControls | Lets you zoom and rotate the view with the mouse
System prompt | Tells the model to return one self-contained HTML file

Example: You set the card's system prompt and ask the model to build a simple object. It writes one HTML file. You save it, open it in a browser, and turn the object with the right mouse button.

For newcomers and experts: A good first try is a simple object, so the answer stays short and easy to check. Open the file in a browser before you ask for changes.

For experts, the author trained the model with QLoRA over 3 epochs. The bf16 type was not supported, and a full fine-tune in fp32 would have been too slow. The datasets are described as good quality but small. The prompt asks for inline JavaScript and CSS, Three.js and OrbitControls from a CDN at a pinned version, a white background and no extra interface elements. The author says a later checkpoint is planned with fixes and wider 3D topics.

You set the system prompt and describe an object
The model writes one HTML file with Three.js
You save the file and open it in a browser
Scroll to zoom and right-drag to rotate

Good to know:

Title: Model facts
Item | Value
Parameters | 4.2B
Weights in BF16 | about 8.4 GB
Weights at 4-bit | about 2.1 GB, an estimate
Formats | Safetensors and GGUF
Built on | Qwen3.5-4B
Language | English

• Long outputs: the model writes a lot, so set a high max_tokens limit.
• Loops: it can get stuck while naming variables. If that happens, stop the run and ask it to move on to the building part.
• Mistakes: it can slip on old library versions or steps in the wrong order. Point them out and it can fix them.
• Training: the datasets are small, and the author used QLoRA.
• Card language: the author says the card was machine translated, so some parts may be unclear.
• License: apache-2.0
• Cost to run: no API key or paid service named.
• Data it reaches: it downloads files from Hugging Face. The card's prompt has the page load Three.js from a CDN, so opening it probably needs internet access.
• Not for: precise or highly detailed models.

On Hugging Face, 13 November 2026

• Task: text generation
• Parameters: 4.2B
• Weights: about 8.4 GB in BF16; about 2.1 GB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 1,735
• Likes: 5
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 1
• Tags: safetensors, gguf, qwen3_5_text, 3d, html, text-generation, conversational

Links

• Model: https://huggingface.co/mondk/Qwen3.5-4B-3D-Html
• Files: https://huggingface.co/mondk/Qwen3.5-4B-3D-Html/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download mondk/Qwen3.5-4B-3D-Html

Then follow the usage in its model card. Nothing in this kit runs by itself.