CCORE GLB-3D-Coder 4B final model and full weights are out
The CCORE GLB-3D-Coder 4B final model and its full weights have been released. This model is a fine-tune of Qwen3.5-4B, designed to generate reusable 3D assets. Users can describe desired assets, and the model will produce a complete Three.js HTML scene with an "Export GLB" button. The generated HTML can be opened in a browser to inspect results and export GLB files for use in tools like Godot, Blender, or Unity.
Unlike other code generation models, this release specifically targets 3D asset creation by generating Three.js HTML scenes with an exportable GLB button, moving beyond simple code snippets.
Time & source
Times shown in UTC
Display time zone: UTC
Local time zone unavailable; showing UTC.
PublishedOffset at this time: UTC+0Nov 5, 2026, 06:14 UTC
IngestedOffset at this time: UTC+0Nov 5, 2026, 07:00 UTC
- Published
- Nov 5, 2026, 06:14
- Ingested
- Nov 5, 2026, 07:00
- Source type
- Dev community
- Tier
- Community
- Source status
- Healthy
Tier is a per-source editorial setting, not a per-item score.
Discussion trend
The percentage is based on collected discussion signal, not new comments or independent people. The curve only compares the same topic across time.
Following up on my previous post : the final version of CCORE GLB-3D-Coder is now available, including the full Safetensors weights.
It’s a full fine-tune of Qwen3.5-4B for generating reusable 3D assets. Describe what you want, and it writes a complete Three.js HTML scene with an Export GLB button. You can open the HTML in your browser, inspect the result, and export it into Godot, Blender, Unity, or other tools that support GLB.
The video shows examples from the final model, including selected side-by-side comparisons with DeepSeek V4.1 Flash. I’m particularly happy with the character designs and the amount of detail it can produce at just 4B parameters.
It covers more than fantasy characters: buildings, environments, weapons, vehicles, laboratory equipment, sports props, furniture, and other objects. It can also generate animations and include them in the exported GLB.
Some generations still need a retry or a small code correction, but I’m very happy with the improvement over the earlier release.
If you’re interested in the next models, or would like to help support future datasets and training runs, there’s more information on the Hugging Face page. Testing it and sharing your results or failure cases also helps a lot.
I made RIFTWARD, a small browser arena using characters, terrain and props generated locally by CCORE GLB-3D-Coder 4B. An orchestration and validation harness helped assemble, refine and integrate the assets and build the gameplay. Play it here: https://huggingface.co/spaces/ccore/RIFTWARD
Feel free to share what you make with it! I’d be really happy to hear that the model has been useful to you :)