Qwen3.5-4B-GLB-3D-Coder

Qwen3.5-4B-GLB-3D-Coder (branded CCORE GLB-3D-Coder) is a 4B-parameter text-generation model fine-tuned from Qwen/Qwen3.5-4B by the community group CCORE. Given a text description of a character, prop, building or environment, it writes a complete standalone Three.js HTML/JavaScript page with a preview and an "Export GLB" button.

The exported GLB files can be opened in Godot or any other editor that supports GLB. The model was trained on the CCORE ThreeJS-3D-GLB-Assets dataset, which has 3,918 tested and rendered examples across 40 domains with validated GLB exports. The model card shows comparisons against DeepSeek V4.1 Flash on prompts outside the training set, and showcase projects such as the RIFTWARD browser game and a community 3D RPG demo.

Weights are published on Hugging Face as Safetensors and as GGUF quantizations (Q4_K_M, Q8_0), so it can run locally in tools such as LM Studio. The card says weights uploaded before November 5, 2026 should be deleted and re-downloaded, because the GGUF filenames were not changed. Output is procedural, low-poly geometry, and quality claims come from the authors' own showcase.