CCORE GLB-3D-Coder 4B Final Weights

Introduction
CCORE just shipped the final checkpoint of GLB-3D-Coder 4B on Hugging Face at ccore/Qwen3.5-4B-GLB-3D-Coder (card dated November 5, 2026; Safetensors push ~05:42 UTC that day). This is the soft-watched follow-through after the October preview: a Qwen3.5-4B full fine-tune that turns an English asset brief into a standalone Three.js HTML page with a working Export GLB button, then drops the mesh into Godot, Blender, Unity, or any GLB-friendly DCC. Same-day companion: the training set ccore/ThreeJS-3D-GLB-Assets went public 2026-11-06 ~00:25 UTC (3,918 validated examples across 40 domains).
Unlike diffusion image-to-3D stacks (TRELLIS, Idol, Meshy), GLB-3D-Coder is a coding model for game props. You edit the generated Three.js (or re-prompt) instead of regenerating a baked mesh from scratch — the maker pitch on the card is “3D you edit, not 3D you regenerate.”
What shipped
- Open weights — Apache-2.0:
Qwen3.5-4B-GLB-3D-Coder-Q4_K_M.gguf(2.71 GB),…-Q8_0.gguf(4.48 GB), plus five Safetensors shards (~17.5 GB original checkpoint) - Optional vision —
mmproj-Qwen3.5-4B-BF16.gguf(~0.68 GB) if your GGUF runtime can take an image reference - Modality — text (English prompts work best) → complete Three.js HTML → browser preview → Export GLB (animation included when you ask for it)
- Open vs API — fully local / self-host; no vendor API required
- Playable proof — RIFTWARD arena showcase on Spaces is assembled from assets generated with this checkpoint (browser playable; no live inference needed to play)
Concrete maker fact: load the Q4 or Q8 GGUF in LM Studio / llama.cpp (Qwen3.5-capable runtime), paste the card’s system prompt (“Create game assets for the Godot editor… Export GLB button…”), set temp 0.8 / top-k 40 / top-p 0.95, leave thinking on, and set max output tokens unlimited — responses average ~20K tokens and some scenes need more. Save the HTML, open it, click Export GLB, import.
Why Titus-lane pipelines should care
Indie and prototype teams still burn hours blocking on greybox props. A 4B local coder that emits editable Three.js + engine-ready GLB is a different lever from Meshy/TRELLIS:
- Engine handoff is the product — GLB lands in Godot / Blender / Unity without a separate remesh/export hop
- Edit the code, not only the prompt — fix a cable route or swap a material in the HTML before export
- Runs on a gaming PC — Q4_K_M is ~2.7 GB; no 80 GB A100 tax like world-model weights
- Dataset is open too — 3,918 tested/rendered GLB examples for fine-tunes or eval harnesses
Caveats from the card (worth keeping honest): some generations need a retry or a small code fix; materials/lighting/animation can diverge between browser preview and the exported GLB. Treat outputs as prototype / greybox / stylized low-poly, then polish in DCC when production art matters.
Where to run it today
- Model: https://huggingface.co/ccore/Qwen3.5-4B-GLB-3D-Coder
- Dataset: https://huggingface.co/datasets/ccore/ThreeJS-3D-GLB-Assets
- Runtimes: LM Studio / llama.cpp (Qwen3.5) for GGUF; transformers-style load for Safetensors
- If you pulled files before Nov 5: re-download — GGUF filenames stayed the same; the Nov 5 checkpoint is the final one
- Not this drop: Comfy partner node, fal/Replicate endpoint, diffusion mesh bake, or a hosted chat UI from CCORE
Same-day takeaway
Preview → final open weights + training set in under 48 hours. Soft watch from the October 3D-Coder preview is cleared: Q4/Q8/Safetensors are on the Hub, RIFTWARD shows a playable assembly path, and the recipe is local-first for Godot greys. Keep Meshy/TRELLIS/Idol for photo-mesh fidelity; reach for GLB-3D-Coder when you want a promptable, editable, GLB-exporting prop pipeline on a small GPU.
—Titus
