The open-source Qwen3.5-4B-3D-Coder model has been released, bringing 3D scene coding to a 4-billion-parameter model, with the ability to generate Three.js code and 3D visuals from text or images. This community-fine-tuned model is targeted at indie developers and creative coders, providing a new tool for interactive 3D scene generation.
What Is Qwen3.5-4B-3D-Coder
Qwen3.5-4B-3D-Coder is an independent fine-tune of Qwen/Qwen3.5-4B, focused on creating interactive 3D scenes with Three.js. The model is a 4 billion-parameter language model, released in both FP16 and quantized GGUF formats. It supports multimodal input, allowing users to generate 3D scenes from text or images, and can be run locally with llama.cpp and LM Studio.
How It Works
The Q4_K_M and Q8_0 versions of the model can accept image inputs when loaded with the original Qwen3.5-4B vision encoder/projector. A 644 MiB BF16 vision projector file is required for image-guided generation. The model generates scenes as self-contained HTML files using the Three.js core ES module from cdnjs. Initial tests have shown substantial improvement in one-shot success and overall scene quality, although occasional duplicate variable name bugs remain.
Performance And Dataset
The training dataset, “CCORE Three.js 3D Scenes,” contains 4,146 examples, including prompts, full thinking traces, and complete HTML/JavaScript responses. The dataset is released under the MIT License.
Use Cases And Benefits
Qwen3.5-4B-3D-Coder can be used by indie developers and creative coders to generate interactive 3D scenes for various applications, such as games, simulations, and visualizations. The model’s ability to generate 3D scenes from text or images makes it a powerful tool for those who want to create complex 3D environments without extensive coding knowledge.
Future Roadmap And Community Involvement
The author plans to release a GLB-3D-Coder variant for reusable 3D assets, a larger 100,000-example dataset, and longer context lengths up to 128,000 tokens. The community is invited to contribute to the project by funding further GPU compute and dataset expansion via cryptocurrency donations. The author’s goal is to keep sharing models, weights, datasets, and improvements openly with the community, making these tools accessible to everyone. Contributions can be made via Ethereum, Solana, or other cryptocurrencies, with every contribution helping to fund more GPU compute, better datasets, new training runs, and future public releases.
Sources
Photo: Matthew Henry / stocksnap (CC CC0 1.0)
