CCORE GLB-3D-Coder 1.7B

Final checkpoint for the 1.7B GLB experiment

The model weights were last updated on November 13, 2026 at 00:15:34 UTC. If you downloaded a GGUF before this update, delete the local copy and download the updated file.

The main tests used F16 and Q8_0. In testing, Q4_K_M produced substantially lower-quality outputs.

Generate Three.js game assets and export them as GLB on modest hardware.

I fine-tuned Qwen3-1.7B on the CCORE GLB dataset to write complete HTML pages with an interactive 3D preview and a GLB export button.

Why Qwen3-1.7B?

I wanted to see how far a small model could go on this task while keeping full causal attention with grouped-query attention across all 28 layers. The goal is a short iteration loop for people with limited VRAM or no dedicated GPU. The Q8_0 file is about 1.83 GB and is my suggested starting point for CPU or low-VRAM use. Generation speed depends on the hardware, runtime, context length, and how much code the model writes.

Examples

Frames from the showcase. The prompts below are suggestions for similar assets.

Weights and release status

All three GGUFs come from the same checkpoint. The FP16 Safetensors weights are also available for Transformers, along with the model configuration and tokenizer.

Quick start

1. Load the GGUF in a Qwen3-compatible runtime such as llama.cpp or LM Studio.
2. Use the system prompt below and describe the asset you want.
3. Save the generated HTML to a file and open it in a browser.
4. Click Export GLB and open the downloaded asset in your 3D editor.

Recommended sampling

System prompt

Create game assets for the Godot editor, which uses GLB files. Create whatever the user requests as a good-quality low-poly asset using only Three.js in one complete standalone HTML file. The HTML must show the asset and include an Export GLB button that downloads a working GLB. If the requested asset needs animation, add it and include the animation in the exported GLB; otherwise make it static.

Training domains

40 domains represented in the GLB dataset

• Characters and creatures — Characters and NPCs, civilian characters, fantasy adventurers, modern specialists, companions and creatures, creatures and companions, enemies, enemies and bosses, bosses, robots and constructs.
• Items and props — Weapons and armor, weapons/armor/equipment, items/tools/pickups, props and furniture, furniture and interiors, household appliances, electronics and optics, musical instruments, workshop and repair.
• Environments and structures — Nature and environment, nature and garden, modular architecture, battlefield scenery, underwater structures, camping and travel, farming and homestead, urban services.
• Vehicles and machines — Ground transport, air and space transport, watercraft, industrial machinery.
• Activities and interactive assets — Cuisine and markets, festivals and ceremonies, sports and playgrounds, fitness and rehabilitation, laboratory and medical, museums and field research, interactive objects, traps and hazards, puzzles and gameplay markers.

Current limitations

In my tests, the model usually produces a 3D asset. Some attempts fail to render or export as GLB and may need a retry.

Support

Future model and dataset releases depend on community support. If you find this project useful, your contribution helps me cover GPU costs and keep those releases going. Support the project on Ko-fi.

License

Apache 2.0. Based on Qwen3-1.7B by the Qwen team.