Back home

CCORE GLB-3D-Coder

hf 0.62
Visit source

Intro

CCORE GLB-3D-Coder is a specialized 4B coding model that turns text into editable 3D assets, published on Hugging Face by CCORE. Model page: https://huggingface.co/ccore/Qwen3.5-4B-GLB-3D-Coder .

How it works: describe a character, prop, building or environment and the model writes a complete Three.js HTML/JavaScript page with a live preview and an Export GLB button; inspect, then export the asset into Godot or any GLB-capable editor.

Highlights:

  • Base: Qwen/Qwen3.5-4B, Apache 2.0;
  • 2026-11-05 release weights: Q4_K_M and Q8_0 GGUF plus original Safetensors; users who downloaded before Nov 5 are told to re-download;
  • Training dataset planned for release in the next few days;
  • Showcase RIFTWARD: Guardians of the Ancient, a playable browser arena on HF Spaces built from assets generated by the model; community dev Naoki Kishida also built a 3D RPG demo with the preview;
  • About 176 likes and ~12k downloads at scan time, newly in HF trending models.

Explain

Why it's interesting

Most 3D generation uses diffusion/mesh reconstruction with hard-to-edit outputs. GLB-3D-Coder takes a "code as asset" route: procedural Three.js builds low-poly models that are readable, editable, reproducible and tiny — great for indie game prototyping.

Small model, vertical task

At 4B parameters it runs locally in LM Studio, llama.cpp, etc.; paired with a coding agent (a starter prompt is provided) it can mass-produce assets and assemble playable games.

Limits

  • Low-poly/procedural style, not photoreal high-poly;
  • Community fine-tune; quality evidence is mostly author showcases;
  • Preview and release weights share filenames — re-download carefully.

Usage

Who it's for

  • Indie game devs and Godot/Three.js users needing fast prototype assets;
  • Creators wanting local small-model 3D generation;
  • Researchers exploring code-generated 3D.

Getting started

  1. Download Q4_K_M or Q8_0 GGUF from https://huggingface.co/ccore/Qwen3.5-4B-GLB-3D-Coder ;
  2. Load in LM Studio / llama.cpp and describe an asset per the model card's prompt ideas;
  3. Open the generated HTML in a browser, preview, click Export GLB;
  4. Try the HF Space ccore/RIFTWARD and its agent starter prompt to build your own game.

Advertisement