If you already run ComfyUI for image models, update to a recent release before adding a 3D model — older builds may miss loaders used by newer text-to-3D checkpoints. From your ComfyUI folder:
git pull then pip install -r requirements.txt
If you are starting from zero, grab the official ComfyUI portable build for your OS from its GitHub releases page and run it once to confirm the UI loads at 127.0.0.1:8188.
Download only from the official Hugging Face repos:
| Variant | Hugging Face repo | Use when |
|---|---|---|
| 4B base (Three.js) | ccore/Qwen3.5-4B-3D-Coder | You want interactive web scenes |
| 4B GLB | ccore/Qwen3.5-4B-GLB-3D-Coder | You want GLB files for Blender/Unity |
| 1.7B GLB | ccore account, 1.7B GLB Coder | You have a smaller GPU |
Place the weights in your ComfyUI models directory following the repo README's layout (the exact subfolder depends on which loader nodes your ComfyUI version provides for this model class).
Community reports suggest the 4B model is comfortable near 9 GB VRAM with standard precision, and workable below that with quantized weights at some quality cost. If you are on 8 GB, prefer the 1.7B GLB variant or a quantized build — and close other GPU-heavy apps first. Exact numbers vary by backend and driver; treat any figure you read (including here) as a starting point, not a promise.
No — you can run the model directly with transformers-style loaders from the Hugging Face repo. ComfyUI is the popular choice because it gives you a node graph, previews and reusable workflows.
Only from the official Hugging Face repos (ccore account: Qwen3.5-4B-3D-Coder and its GLB/1.7B variants). Avoid re-uploads — weights from unofficial mirrors can be modified or outdated.