Qwen3.5-4B-3D-Coder

Qwen3.5-4B-3D-Coder

Writes code and interactive Three.js 3D scenes from text prompts, and from a reference image when you add the optional vision file. An early preview of a 4B fine-tune for people testing small models.

Made by ccore 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-Coder is a small coding model tuned to write interactive 3D scenes with Three.js, and to code in general. It is an independent fine-tune of Qwen3.5-4B by CCORE, not an official Qwen release. The author calls it an early preview of a model still in training.

Who it suits: It suits developers and hobbyists who want to test how far a small local model can go at coding and 3D scene generation, and who can run a GGUF file or Transformers.

Industry | Who uses it | How they use it | Example
Web development | Front-end developers | Generate Three.js scenes as a starting point for a page | Ask for a scene in plain words and run the code that comes back
Games and 3D | Independent creators | Sketch a voxel scene from a text prompt or a reference image | Attach a picture of a voxel tree and ask for the same scene
Education | Teachers and students | Show how a short prompt turns into a running 3D scene | Ask for a small scene and read through the code it writes
Research | Model tinkerers | Study what a 4B fine-tune can do with 3D code | Compare scenes across the checkpoints or use the public dataset

People who need dependable output on the first try can skip it for now, since some scenes still need another attempt or a fix.

Where to use it:

• llama.cpp: the llama-server WebUI loads the GGUF files and accepts an attached image
• Hugging Face Transformers: the FP16 weights come in Safetensors format
• Other frontends for GGUF files: image input depends on their own multimodal support
• A web browser: where the Three.js scenes it writes can run

Use case | Who | What they get back
Text to 3D scene | Developers | A complete HTML and JavaScript Three.js scene
Image to 3D scene | Creators | A Three.js scene that reads a reference picture without copying every detail
General coding | Programmers | Code answers for everyday programming questions

How to set it up: You need these first.

• llama.cpp with llama-server, where the author tested build b11889
• One GGUF file from the model page, either Q4_K_M or Q8_0
• For image input, the BF16 vision projector file, about 644 MiB

1. Download one GGUF file from the model page, either Q4_K_M or Q8_0.
2. If you want image input, download the BF16 vision projector too. One file works with both versions.
3. Start llama-server with the model and the projector, using the command below.
4. Open the server's WebUI, attach an image if you like, and send your prompt.

llama-server \
  -m Qwen3.5-4B-3D-Coder-Q4_K_M.gguf \
  --mmproj mmproj-Qwen3.5-4B-BF16.gguf \
  --n-gpu-layers all --flash-attn on \
  --ctx-size 65536 --parallel 1

With the Q8_0 version, swap the Q4_K_M filename after -m for the Q8_0 one. Adding these two flags can shrink the memory the KV cache takes. The author tried them with vision at a configured 64K context.

--cache-type-k q8_0 --cache-type-v q8_0

The FP16 weights are published too. In the author's check they loaded and generated text with Transformers 5.20.2.

How to use it: You chat with the model in the llama-server WebUI. Use the Three.js system prompt from the model page, then ask for a scene in plain words. With an image attached, the author's example prompt was this.

create this scene using voxels 3d

The model returns the scene as code. For the FP16 weights the author sets temperature to 0.6, top-k to 20, top-p to 0.95 and the repetition penalty to 1.0, and leaves caching on. The model page lists the settings for the GGUF files.

Example: You attach a reference picture of a voxel tree, add the Three.js system prompt and send the prompt above. The model returns a Three.js scene called Tree Island. It follows the picture but does not reproduce every detail.

For newcomers and experts: It is a small model that writes code, and its special skill is turning a short request into a 3D scene that runs in a browser. The first small thing to try is one simple Three.js scene with the system prompt on, then running the code it gives you.

It was fine-tuned from Qwen3.5-4B on 4,146 scene examples that hold prompts, full thinking traces, complete HTML and JavaScript answers and complexity levels. The weights come as Q4_K_M and Q8_0 GGUF files and as FP16 Safetensors in four shards. Image input pairs a GGUF file with the unmodified BF16 projector from the original Qwen3.5-4B. The projector adds to memory use. Pictures you attach also take up part of the context window, which your prompt and the generated code fill too.

You write a prompt, with an optional reference image
llama-server loads the GGUF file and the projector
The model reasons about the scene
It writes a full HTML and JavaScript Three.js scene
You run the code in a browser

Good to know:

• It is an early preview and an independent fine-tune, not an official Qwen release
• Some scenes still need another attempt or a correction
• The author checked the vision setup on their own machine with llama.cpp b11889, using brief runs rather than a test at the full context size
• If you downloaded the files before the October 27 checkpoint, download them again
• The model is under the Apache-2.0 license, and the training dataset is under the MIT License
• A technical report for the training experiment is available as a PDF on the model page

The public training dataset holds these example counts.

Title: Training dataset size
Training examples: 3933 examples
Validation examples: 213 examples

On Hugging Face, 10 November 2026

• Parameters: 4.5B
• Weights: about 9.1 GB in F16; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 289,466
• Likes: 958
• License: apache-2.0
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 16
• Tags: safetensors, gguf, qwen3_5, code, threejs, 3d, preview, conversational

Links

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

Get the model

Download its files with Hugging Face's command line:

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

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