CCORE's Qwen3.5-4B-3D-Coder Turns Text Prompts Into 3D Browser Scenes
An independent 4B fine-tune of Qwen3.5 turns plain text prompts into self-contained Three.js scenes, with weights, dataset, and GGUFs public.
- Qwen3.5-4B-3D-Coder is an independent 4B fine-tune generating Three.js scenes as single HTML files.
- Ships as GGUF: Q4_K_M (2.71 GB) and Q8_0 (4.48 GB), runnable in llama.cpp, Ollama, LM Studio.
- Training dataset of 4,146 scenes with reasoning traces released under MIT license.
- Full FP16 weights, technical report, and additional quantizations planned for release.
- Recommended settings: temperature 0.6-0.8, top-p 0.9-0.95, top-k 20-40; disable MTP in LM Studio.
- Known bugs are minor, mostly duplicate variable names rather than structural failures.
Independent developer CCORE has released an early preview of Qwen3.5-4B-3D-Coder, a 4-billion-parameter fine-tune that generates interactive Three.js scenes from text prompts. The model returns a self-contained HTML document containing the JavaScript needed to render the scene in a browser.
At publication, the model had nearly 7,000 downloads and ranked near the top of Hugging Face’s trending list while training remained in progress. Those counters are volatile, and scene quality still requires direct evaluation. The practical draw for developers is a small, locally runnable model with a tightly defined output contract.
| Category | Published detail |
|---|---|
| Publisher | CCORE, an independent developer |
| Status | Early preview; training is ongoing |
| Base model | Qwen/Qwen3.5-4B |
| Output | One HTML file with Three.js and JavaScript |
| Library contract | Three.js 0.160.0 core ES module from cdnjs |
| Published quantizations | Q4_K_M at 2.71 GB and Q8_0 at 4.48 GB |
| Training corpus | 4,146 examples: 3,933 training and 213 validation |
| Licensing | Dataset under MIT; consult the model card for weight terms |
A tight contract shrinks the problem
Qwen3.5-4B is a general-purpose model from Alibaba’s Qwen project. CCORE’s release is an independent derivative unaffiliated with official Qwen releases, and it concentrates on a narrow task: translating a spatial description into browser-renderable Three.js code.
The fixed output format reduces variation across training examples. Every response targets one document, one library version, one import style, and inline implementations for features such as camera controls. That consistency gives a 4B model fewer APIs and build configurations to manage than a general coding assistant encounters.
The released examples cover stylized architecture, gardens, voxel environments, furniture, workshops, mechanical objects, and animated infrastructure. Several prompts request visible cubic blocks, clear silhouettes, depth, and readable spatial relationships, making voxel scenes a recurring part of the training distribution.
This story is for Pro members
You've reached the end of the free preview. Upgrade to AlphaSignal Pro to read the full article - and everything else behind the paywall.
Comments
