CCORE GLB-3D-Coder

CCORE GLB-3D-Coder

Writes a Three.js page from a text prompt that previews a low-poly 3D asset and exports it as GLB for Godot or other editors. Made for game developers and 3D hobbyists.

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: CCORE GLB-3D-Coder is a 4B text generation model from ccore, trained from the Qwen3.5-4B base to write code for 3D assets. You describe a character, prop, building or scene in plain words. It replies with a full Three.js page that shows a preview and has an Export GLB button, so the asset can move into a game editor.

Who it suits: It suits people who make low-poly game assets and want a quick first version of a character, prop or environment. It also fits anyone who would rather generate a scene from code than model it by hand.

Industry | Who uses it | How they use it | Example
Game development | Indie developers and game jam teams | Generate low-poly assets and export them as GLB files | A frozen goblin watch camp for a level
Game prototyping | Designers testing an idea | Build a rough scene to judge a concept before art work starts | A bioluminescent cave tram station
Education | Teachers of 3D graphics and web code | Show students how code turns into a 3D shape | A coral-armored lighthouse keeper
Web 3D | Front-end developers | Get a Three.js page they can read and edit | A character whose colors they change in the code

You can skip it if you need highly detailed or realistic models, because it is built for simple low-poly shapes and procedural scenes.

Where to use it: The model runs on your own machine, and its output opens in a browser.

• LM Studio or any other app that loads GGUF model files
• A browser, where the generated Three.js page runs and shows the preview
• Godot, or any other editor that opens GLB files
• A coding agent that talks to a local model API, which the RIFTWARD game page describes

Use case | Who | What they get back
Characters and props | Game artists | A Three.js page and a GLB file of the character
Buildings and environments | Level designers | A scene preview they can export for their editor
Assets for a browser game | Web game developers | GLB files that load in the game without live model inference

How to set it up: You pick a weight format, load it in a local model app and prompt it.

You need:

• A local model app such as LM Studio
• One set of weights, in Q4_K_M, Q8_0 or Safetensors form
• A browser to open the page the model writes
• Godot or another GLB editor if you want to bring the asset in

1. Choose a weight format on the model page.
2. Download it into LM Studio or the app you use for local models.
3. If your copy is from before November 5, 2026, remove it and fetch the weights fresh. The file names stayed the same, so check that you load the new ones.
4. Load the updated weights and start a chat with the model.

How to use it: Type a description of the asset you want. The model answers with one HTML page of Three.js code. Open that page in a browser, check the preview, then press Export GLB to get the file. A plain request looks like this:

Create a coral-armored lighthouse keeper.

Longer prompts work too. The project's own tests ran from short requests to detailed descriptions that name clothing, tools and proportions.

Example: Ask for a camp where goblins keep watch in the snow. The model writes a Three.js page, and opening it shows the camp in the preview. If it looks right, you press Export GLB and load the file in Godot.

For newcomers and experts: Think of it as a helper that draws in code. You type what you want to see, it writes a small web page, and the page shows the 3D result. Start with one character or prop, open the page it returns and press Export GLB.

Underneath, it is a 4B model that starts from Qwen3.5-4B. Its released training set holds 3,918 examples from 40 domains, each one tested, rendered and paired with a checked GLB export. It writes geometry as JavaScript instead of producing a mesh directly, so you can read and edit the result. The trade-off is size, and the project compares its output against DeepSeek V4.1 Flash 1017 on the same prompts.

You describe the asset in a prompt
The model writes a Three.js page
You open it and check the preview
Export GLB saves the asset file
Godot or another editor loads it

Good to know:

• Size and base: a 4B model trained from Qwen3.5-4B
• Weights come as Q4_K_M, Q8_0 or Safetensors
• Dataset: CCORE ThreeJS-3D-GLB-Assets, with 3,918 examples from 40 domains
• Copies from before November 5, 2026 need a fresh download
• RIFTWARD, a browser arena game, uses assets from this model and plays without live model inference
• License: Apache 2.0

On Hugging Face, 10 November 2026

• Task: text generation
• Parameters: 4.5B
• Weights: about 18.2 GB in F32; about 2.3 GB at 4-bit (estimate)
• Formats: Safetensors, GGUF
• Downloads in the last 30 days: 38,696
• Likes: 289
• License: apache-2.0
• Languages: en
• Based on: Qwen/Qwen3.5-4B
• Quantized versions: 9
• Fine-tunes: 1
• Tags: safetensors, gguf, qwen3_5, code, threejs, 3d, godot, text-generation, conversational

Links

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

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