A 2.71GB local AI model can now generate interactive 3D scenes from just a text prompt 🤯

Meet Qwen3.5-4B-3D-Coder, a new open-source community fine-tune by CCORE, built for coding and 3D scene generation using Three.js.

And the interesting part is that it doesn't just generate static images. It writes HTML + JavaScript to create actual 3D environments that you can open and interact with in your browser.

Some of the developer's examples include:

🏯 Japanese gardens and temples
🚀 Voxel rocket launch sites
🐉 Fantasy landscapes and architecture
🪑 Furniture and workshop interiors
⚙️ Even detailed mechanical objects like CPU sockets

All generated using a model with just 4B parameters.

Here's what makes it even more interesting:

🧠 Model: 4B parameters
💾 Q4_K_M: 2.71GB
⚡ Q8_0: 4.48GB
💻 Runs locally with Ollama, llama.cpp and LM Studio
🔓 Apache 2.0 licensed

No cloud API is required for model inference.

And this is still an early preview. Training is expected to finish on October 28, with full weights, additional quantizations and the training dataset planned for release afterwards.

There are still occasional code errors, and its performance hasn't been independently benchmarked yet.

But seeing smaller models become specialized enough to generate interactive 3D environments is genuinely interesting.

Imagine using something like this to prototype game environments, interactive websites or 3D concepts directly on your own computer.

🔗 Model + examples:
https://huggingface.co/ccore/Qwen3.5-4B-3D-Coder

What would you build with a small local model like this? 👀