Latest in AI
GLB-3D-Coder Preview: A Small Qwen-Based Model for Godot Assets
CCORE's updated checkpoint writes editable Three.js scenes, but exporting a file is only one part of making a usable game asset.
Independent developer CCORE updated GLB-3D-Coder on November 2 with new downloadable checkpoints and optional image-reference support. The Qwen3.5-4B-based preview generates Three.js code that can export 3D assets for Godot. It is a small-model experiment with a practical output format, not an official Qwen release or a finished asset-production system.
What GLB-3D-Coder produces
The model card describes a workflow in which the model writes HTML and JavaScript, displays a scene in the browser, and includes an export button for a GLB file. CCORE shows examples imported into Godot, while warning that some generations need another attempt or a code correction.
That makes the code an intermediate representation. A developer can inspect how the scene is assembled rather than receiving only an opaque final asset. The tradeoff is that a plausible-looking scene can still contain broken code or an export that behaves differently in the destination editor.
What changed in the November 2 checkpoint
CCORE dates the update to 04:50 UTC. Both quantized checkpoints were replaced without changing their filenames. The author also added an optional vision projector for image inputs in compatible runtimes. Training continues; full Safetensors weights and the training dataset are still planned rather than released.
The repository files show the Q4_K_M download at 2.71 GB and Q8_0 at 4.48 GB, plus a separate 676 MB projector. Those are file sizes, not total memory requirements. The repository carries an Apache-2.0 license label.
Unchanged filenames make version tracking particularly important. Two people comparing results may have downloaded different checkpoints even when their local model names match. Recording the repository revision alongside a test is more useful than recording the filename alone.
Where the preview fits
For developers prototyping game props or experimenting with procedural scenes, the interesting question is whether this shortens the route from an idea to an editable starting point. It does not need to replace an artist's whole workflow to be useful there.
A successful export is not the same as a production-ready asset. Geometry, materials, animation, visual consistency and runtime performance still need checking inside the intended game. The author's examples demonstrate selected outputs, not a measured success rate across arbitrary requests.
The useful next evidence would be repeatable comparisons across prompts and checkpoints, including failures after import. That would show whether specialization is improving the complete workflow rather than only making the browser preview more convincing.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help technical founders explain model releases and developer workflows in ways people can understand and share.