Qwen 3.5 4B 3D Coder is an open-weight, 4-billion-parameter language model fine-tuned for one job: turning a text description into working 3D code. Its headline trick is generating interactive Three.js scenes — you describe a layout or object, the model writes the scene code, and you open it in a browser. A companion variant, Qwen 3.5 4B GLB 3D Coder, targets the GLB file format instead, which is the interchange format understood natively by Blender, Unity, Unreal and most 3D tools.
The model was published on Hugging Face by the ccore account in late October 2026 and has since accumulated hundreds of thousands of downloads, with an active stream of community write-ups on setup and VRAM use. A smaller 1.7B GLB variant exists for machines that can't feed the 4B model.
| Variant | Output | Best for |
|---|---|---|
| Qwen 3.5 4B 3D Coder | Three.js scene code (interactive web) | Web experiences, quick 3D previews, landing-page hero scenes |
| Qwen 3.5 4B GLB 3D Coder | GLB files (binary glTF) | Game assets, Blender editing, Unity/Unreal pipelines, 3D printing prep |
| Qwen 3 1.7B GLB Coder | GLB files (smaller model) | Lower-VRAM machines, batch prototyping |
Choose by destination: if the result lives in a browser, the Three.js variant speaks your language natively. If the result must be edited or imported elsewhere, use the GLB variant and read our GLB export pipeline guide.
ccore account. Always download from the official repos.It is an open-weight 4-billion-parameter model, published on Hugging Face by the ccore account, that generates interactive Three.js 3D scenes from text prompts. A GLB-output variant targets direct use in 3D software.
The model weights are published openly on Hugging Face, so local generation is free. Check the specific repo's licence file for commercial-use terms before shipping outputs in a product.
Community reports put the 4B variant at roughly 9 GB of VRAM for comfortable use; a smaller 1.7B GLB variant exists for lower-end GPUs. See our VRAM requirements page for details.