Qwen3.5-4B 3D Coder and GLB 3D Coder: Make 3D Scenes and Godot Game Assets Locally From a Prompt

What if a model small enough for a laptop could turn “a frozen goblin watch camp” into a 3D scene you can spin around in your browser, then export it straight into Godot? That is the idea behind two open models from independent developer CCORE: Qwen3.5-4B-3D-Coder and the newer Qwen3.5-4B-GLB-3D-Coder.

Both are 4B fine-tunes of Qwen3.5-4B under Apache 2.0. Instead of generating a mesh directly, they write a complete Three.js web page that builds the object in code. The GLB version, released on 5 November 2026, adds an Export GLB button so the result drops into a game engine.

The first model has passed 220,000 downloads in a month and is trending on Hugging Face. This guide covers what each model does, exact files and sizes, how to run them with Ollama, llama.cpp or LM Studio, the right settings, Godot and Blender import, limits and fixes.

What Are the 3D Coder Models?

Most text-to-3D tools generate a mesh or a Gaussian splat with a diffusion model. CCORE’s models take a different route: they are language models that write code. You describe an object, and the model replies with one HTML file that uses the Three.js library to build, light and animate it.

Open that file in any browser and you get a live, rotatable 3D scene. Because it is code, you can read it, tweak colours or sizes, and run it offline.

CCORE calls the project “an experiment exploring the feasibility of small language models” for 3D work. AI Weekly covered the first release on 28 October 2026, describing CCORE as a solo developer.

Note: the model card is clear that this is “not an official Qwen release”. It is an independent fine-tune built on Alibaba’s Apache 2.0 Qwen3.5-4B.

3D-Coder vs GLB-3D-Coder

There are two models with different jobs. Pick based on what you want at the end.

If you just want cool scenes to look at or embed on a web page, use 3D-Coder. If you make games, use GLB-3D-Coder.

Specs and Files

All sizes below are from the Hugging Face file lists.

The GLB version also comes in Q4_K_M and Q8_0 GGUF. CCORE used Q8_0 for its recorded evaluation. A LiteRT-LM conversion for phones exists too (litert-community/Qwen3.5-4B-3D-Coder), but its own card says Android is not yet recommended.

Hardware and VRAM

CCORE does not publish VRAM figures, so the numbers below are my estimates from the file sizes plus room for a long context. Treat them as a guide.

Speed matters more than memory here, because answers are long. As simple maths: at 100 tokens per second, a 20,000-token answer takes over three minutes. Our VRAM guide explains how context length adds to memory use.

Run It With Ollama

Ollama is the easiest route. It can pull GGUF files straight from Hugging Face.

1. Install Ollama from ollama.com for Windows, macOS or Linux.
2. Pull and run the model. For game assets: ollama run hf.co/ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M. For scenes: ollama run hf.co/ccore/Qwen3.5-4B-3D-Coder:Q4_K_M.
3. Raise the context. Inside the chat, type /set parameter num_ctx 65536. The default context is far too short for these answers.
4. Remove the output cap. Type /set parameter num_predict -1 so the reply is not cut off mid-HTML.
5. Set the system prompt with /set system followed by the prompt from the model card (see Settings).
6. Ask for an object, then copy the HTML from the reply into a file such as goblin.html and open it in Chrome, Edge or Firefox.

Tip: chatting in the terminal makes copying 20K tokens painful. Point a front end like Open WebUI at Ollama, and copy the code block with one click.

Run It With llama.cpp

llama.cpp gives you the most control and an OpenAI-compatible server at http://localhost:8080/v1.

1. Get a recent llama.cpp build. CCORE tested vision with build b11889, so use that or newer.
2. Start the server with the card’s command: llama serve -hf ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M.
3. Set a long context. The card’s example uses --ctx-size 65536 --parallel 1.
4. Save memory on Q8_0 by adding --cache-type-k q8_0 --cache-type-v q8_0.
5. For image input, download mmproj-Qwen3.5-4B-BF16.gguf and load it with --mmproj. Image tokens share the same context window.
6. Open the built-in web UI at http://localhost:8080, set the system prompt and sampling values, and send your request.

Not sure which runtime to choose? Our vLLM vs llama.cpp vs Ollama comparison covers the trade-offs. vLLM also works for the GLB model: vllm serve "ccore/Qwen3.5-4B-GLB-3D-Coder".

Run It in LM Studio

1. Search for ccore/Qwen3.5-4B-GLB-3D-Coder (or the 3D-Coder) in LM Studio’s model search and download Q4_K_M or Q8_0.
2. Turn off multi-token prediction (MTP). LM Studio enables it by default, and the card warns the model may not work properly with it on.
3. Set the context length to 32K-64K and the max response length to unlimited.
4. Paste the system prompt in the system prompt box and set the sampling values below.

Warning: MTP is the most common reason the model “doesn’t work” in LM Studio. Switch it off before you blame the model.

Our LM Studio guide shows where these settings live in the current app.

System Prompt and Settings

These models were trained with a fixed system prompt, so use it exactly as given on each model card. Here is what each one says in short.

3D-Coder system prompt (key rules)

• “Use only the Three.js core ES module from https://cdnjs.cloudflare.com/ajax/libs/three.js/0.160.0/three.module.min.js.”
• “Do not import addons or legacy examples/js paths; implement any camera controls directly inside the HTML.”

GLB-3D-Coder system prompt (key rules)

• “Create game assets for the Godot editor, which uses GLB files.”
• Build a good-quality low-poly asset using only Three.js in one standalone HTML file.
• “The HTML must show the asset and include an Export GLB button that downloads a working GLB.”
• Add animation only when the user asks for it.

Sampling settings

Note: the card says quality with thinking turned off or limited has not been validated. Leave thinking on, even though it makes replies longer.

Export a GLB Into Godot or Blender

Godot 4

1. Open the HTML file the model wrote in your browser and check the asset from a few angles.
2. Click Export GLB. The browser downloads a .glb file.
3. Copy the GLB into your Godot project folder. Godot imports glTF/GLB files automatically and shows them in the FileSystem dock.
4. Drag the file into your scene, or double-click it to open the import settings and create an inherited scene.
5. Check materials and animations. The card warns they can look different in the editor than in the browser preview.

Blender

Use File, Import, glTF 2.0 (.glb/.gltf). From there you can clean up the mesh, add better materials or rig a character. Our best AI for Blender animation guide lists add-ons that help with the next steps.

A community member, Naoki Kishida, has already used GLB-3D-Coder assets in a playable 3D RPG demo, according to the model card.

Prompt Ideas That Work

Short, specific English prompts work best. These come from the model cards:

• GLB: “Create a coral-armored lighthouse keeper.”
• GLB: “Create a frozen goblin watch camp.”
• GLB: “Create a village blacksmith wearing a blue apron.”
• GLB: “Create a white laboratory microscope with a rotating objective turret.”
• 3D-Coder: a canal lock and weir complex with a central island.
• 3D-Coder: an animated air traffic control tower with a turning radar and beacon.

The GLB model was trained on 25 asset categories, including characters and NPCs, enemies, bosses, weapons and armour, props and furniture, modular architecture, vehicles, watercraft, traps and hazards, festival decorations and puzzle markers.

Tip: ask for one asset at a time and say “static” or “animated” up front. Long multi-object scenes are where most retries happen.

How It Compares With Other Text-to-3D Options

The code route gives you clean, light geometry that is easy to tweak, but it will never look photoreal. Diffusion-based 3D looks more realistic but produces heavier, messier meshes.

Limits and Licence

• Retries: the cards say most first attempts work, but some need another try or a small manual fix.
• Common bug: duplicate variable names in the JavaScript, according to the 3D-Coder card.
• Simple prompts only: training data uses simple prompts, so very detailed or unusual requests generalise less well.
• Preview vs editor: materials, lighting and animations can change between the browser and Godot.
• No formal benchmarks: CCORE shares screenshot galleries, not scores. Any quality claims are the developer’s own.
• English works best.
• Licence: both models are Apache 2.0, so commercial use is allowed if you keep the licence and attribution notices. The 3D-Coder dataset is MIT “to the extent of the rights held by the dataset publisher”.

CCORE’s roadmap includes a 100,000-example dataset with 8K-128K token contexts, aimed at more reliable one-shot results and use inside an agent harness.

Troubleshooting

The reply stops halfway through the HTML

Your output limit is too low. Set max tokens to -1 (unlimited) and raise the context to 32K-64K. The GLB card specifically warns that an 8K limit cuts replies off.

The page is blank when I open it

Open the browser’s developer console (F12). A “has already been declared” error means duplicate variable names; rename one, or regenerate. Also check you are online the first time, because the page loads Three.js from cdnjs.

Nonsense output in LM Studio

Turn off multi-token prediction (MTP) in the model settings.

The GLB looks wrong in Godot

Expect some differences in lighting and materials. Re-check the asset in Godot and adjust materials there.

Results got worse after an update

If you downloaded GLB-3D-Coder before 5 November 2026, delete the files and download again. The GGUF filenames did not change, so you may still be loading old weights.

Image input does nothing

Load the mmproj-Qwen3.5-4B-BF16.gguf projector alongside the model, and use a runtime that supports Qwen3.5 multimodal GGUF.

My Take

I like this project because it picks a narrow job and does it with a tiny model. A 2.7 GB file that turns a sentence into an editable, exportable game prop is a real tool for indie devs, teachers and anyone prototyping a game.

The trade-off is time. Long thinking plus 20K-token answers means minutes per asset, and you will retry some. On a mid-range GPU that is fine for a prop library built over an afternoon.

I would start with the GLB model in Ollama at Q8_0 if you have 10 GB of VRAM, since that is what CCORE evaluated. Keep a bigger general model, like Qwen3.8-27B, for fixing tricky code when the small one gets stuck. More local coding picks are in our best Ollama models list.

Frequently Asked Questions

Is Qwen3.5-4B-3D-Coder made by the Qwen team?

No. It is an independent fine-tune by CCORE built on Qwen3.5-4B. The card says it is not an official Qwen release.

Can I use the 3D assets in a commercial game?

The models are Apache 2.0, which allows commercial use with the licence and attribution notices kept. Check the final asset yourself before shipping.

What is the difference between 3D-Coder and GLB-3D-Coder?

3D-Coder makes interactive Three.js scenes for the browser. GLB-3D-Coder makes low-poly game assets with an Export GLB button for Godot.

How much VRAM do I need?

CCORE gives no official figure. Based on the 2.71 GB Q4_K_M file plus a long context, I estimate about 5-8 GB of VRAM at 64K context.

Does it work offline?

The model runs fully offline. The generated HTML loads Three.js from cdnjs, so the page needs internet the first time unless you save a local copy of the library.

Can it build a 3D model from an image?

Yes, with the optional vision projector file loaded. CCORE describes these as short tests, not full stress tests.

Does it work with Unity or Unreal?

It targets Godot, but GLB is a standard format. Unity and Unreal can import glTF/GLB through their own importers or plugins.

Why are the answers so long?

The model thinks before it writes and outputs full HTML and JavaScript. GLB answers average about 20,000 tokens, according to the card.

Sources

• Hugging Face model card: ccore/Qwen3.5-4B-3D-Coder (files, settings, release timeline, licence)
• Hugging Face model card: ccore/Qwen3.5-4B-GLB-3D-Coder (5 Nov 2026 release, system prompt, settings, known issues)
• Hugging Face dataset: ccore/ThreeJS-3D-Scenes (size, fields, teacher model, MIT licence)
• Hugging Face: litert-community/Qwen3.5-4B-3D-Coder (phone conversion and test results)
• AI Weekly: “Solo Dev Opens Qwen3.5-4B-3D-Coder” (28 Oct 2026)
• MindStudio blog: Qwen3.5-4B GLB 3D Coder local guide