Qwen3.5-4B-3D-Coder-GGUF

Qwen3.5-4B-3D-Coder-GGUF

Writes code and builds interactive 3D scenes as self-contained Three.js pages, using GGUF copies of an early-preview Qwen3.5-4B fine-tune. For developers who prototype 3D scenes locally.

Made by prithivMLmods 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: This repository holds GGUF copies of Qwen3.5-4B-3D-Coder, an early-preview fine-tune of Qwen3.5-4B made by ccore. It aims at general coding and at interactive 3D scenes written as self-contained Three.js HTML and JavaScript. Training used an MIT-licensed set of 4,146 Three.js scene examples with prompts, reasoning traces and complete code.

Who it suits: Developers and hobbyists who want a small local model to draft Three.js scenes, and who accept that early-preview output may need fixes.

Industry | Who uses it | How they use it | Example
Web development | Front-end developers | Draft a 3D scene as a single HTML page | Ask for a rotating cube with lighting
Education | Teachers and students | Learn Three.js by editing generated scenes | Read and change the code the model returns
Creative coding | Hobbyists | Prototype interactive 3D ideas on a laptop | Run a small quantized file with no server

Anyone who needs measured accuracy can skip it, since the card gives no benchmark numbers.

Where to use it:

• llama.cpp: the card's own section is about running GGUF files there.
• LM Studio: the card says to disable multi-token prediction.
• A browser, to open the Three.js pages the model writes.
• Windows, macOS or Linux: the download steps run on all three with Python.

Use case | Who | What they get back
3D scene drafts | Developers | A self-contained HTML page with Three.js code
Coding help | Developers | General code from a small 4B-class model
Image-aware prompts | Developers | Vision support through the separate projection file

How to set it up: What you need first:

• Python with pip, to download the files.
• A GGUF runtime, such as llama.cpp or LM Studio.
• Memory for the file you pick, from about 2.3 GB to 8.4 GB.

1. Install the Hugging Face hub package with pip.
2. Download the repository, or just the quant file you want.
3. Load the file in your runtime. For vision, load the projection file too.

pip install -U huggingface_hub
hf download prithivMLmods/Qwen3.5-4B-3D-Coder-GGUF

Title: Which file to pick
Type | Size | Use
BF16 | 8.42 GB | Full-precision weights, best quality, biggest file
Q8_0 | 4.48 GB | Close to the original, rarely needed
Q6_K | 3.46 GB | Near-original quality at a smaller size
Q5_K_M | 3.07 GB | Strong quality, a recommended pick
Q4_K_M | 2.71 GB | The balanced default for most machines
Q3_K_L | 2.42 GB | Lower quality for machines with little memory
mmproj-bf16 | 676 MB | Projection file that adds image input

How to use it: Load a quant file and ask for a scene in plain words, with a system prompt about Three.js if you have one. The model answers with HTML and JavaScript that you save and open in a browser. Retry or fix small errors by hand, because the card warns that some outputs need it.

Example: You ask for a small interactive scene, such as a spinning cube with a light and a ground plane. The model returns a complete HTML page with Three.js code. You save it, open it in a browser and adjust anything that does not run.

For newcomers and experts: A quant is a smaller copy of a model that runs on ordinary computers. A good first step is the Q4_K_M file, which the card calls its default for most uses, followed by one short scene request.

The base is a 4B fine-tune of Qwen3.5 for code that builds 3D scenes, and the card describes it as early preview. Quality drops as the quant gets smaller. The card reports no benchmarks and shows its results through sample scenes instead. A separate repack of the model with multi-token prediction exists, but this repository does not have it.

Pick a quant file that fits your memory
Load it in llama.cpp or LM Studio
Ask for a Three.js scene in plain words
Save the HTML and open it in a browser

Good to know: The card lists the size of each file.

Title: File size by quant type
BF16: 8.42 GB
Q8_0: 4.48 GB
Q6_K: 3.46 GB
Q5_K_M: 3.07 GB
Q4_K_M: 2.71 GB
Q3_K_L: 2.42 GB
Q3_K_M: 2.26 GB

• Status: an early preview, and some outputs may need retries or small code fixes.
• Evidence: no quantitative benchmarks, only sample 3D scenes.
• LM Studio: turn off multi-token prediction.
• License: apache-2.0
• Cost to run: no API key, account or paid service named.
• Data it reaches: files you download from Hugging Face, and the prompts you write.
• Not for: work that needs verified accuracy, or models that must be ready to ship without review.

On Hugging Face, 14 November 2026

• Task: image text to text
• Library: transformers
• Parameters: 4.2B
• Weights: about 2.1 GB at 4-bit (estimate)
• Formats: GGUF
• Downloads in the last 30 days: 9,492
• Likes: 13
• License: apache-2.0
• Languages: en
• Based on: ccore/Qwen3.5-4B-3D-Coder
• Tags: transformers, gguf, llama-cpp, qwen3_5, code, threejs, 3d, preview, image-text-to-text, conversational

Links

• Model: https://huggingface.co/prithivMLmods/Qwen3.5-4B-3D-Coder-GGUF
• Files: https://huggingface.co/prithivMLmods/Qwen3.5-4B-3D-Coder-GGUF/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download prithivMLmods/Qwen3.5-4B-3D-Coder-GGUF

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