🚀 A 4B AI model is learning to build surprisingly detailed 3D worlds using nothing but code. And you can run it locally.

Independent developer CCORE has released an early preview of Qwen3.5-4B-3D-Coder, a specialized fine-tune of Alibaba's Qwen3.5-4B designed for general programming and interactive 3D scene generation.

The interesting part? It generates actual Three.js code, not images or videos.

You describe a scene in natural language, and the model writes the HTML and JavaScript needed to render it in a browser.

Despite having just 4 billion parameters, early examples show surprisingly elaborate environments.

🏗️ What can it generate?

The developer has demonstrated:

▪️ Architectural environments: Japanese temples, traditional houses, villages, and detailed buildings

▪️ Nature scenes: Japanese gardens, bonsai trees, landscapes, and floating environments

▪️ Voxel worlds: Block-based environments with complex spatial layouts

▪️ Detailed objects: Furniture, bicycles, mechanical components, and electronics

▪️ Fantasy environments: Dragonrider settlements, towers, platforms, and interconnected structures

▪️ Animated scenes: Objects with movement, such as rotating radar antennas and flashing lights

▪️ General-purpose code: The fine-tune also targets broader programming tasks rather than exclusively 3D generation

The examples include a sakura temple garden, a sunset village, a voxel launch site, a bookbinding workshop, an Intel LGA 3647 socket, and a dragonrider aerie.

🧠 How does it work?

This isn't a traditional text-to-3D model that predicts meshes directly.

Instead, it writes procedural 3D scenes using Three.js, the JavaScript library for rendering interactive graphics in a browser.

The generated code defines the geometry, materials, lighting, camera, object placement, and scene behavior.

That means the result can be inspected, modified, and extended like a normal web project.

No Blender integration, MCP server, or separate 3D generation service is required for the basic workflow.

The author currently recommends generating self-contained HTML files using the Three.js 0.160.0 core module without additional libraries.

⚡ Small enough for local inference

Two GGUF versions are currently available:

▪️ Q4_K_M: 2.71 GB, 4-bit quantization

▪️ Q8_0: 4.48 GB, 8-bit quantization

▪️ Base model: Qwen3.5-4B

▪️ License: Apache 2.0

You can use compatible local inference tools such as LM Studio, llama.cpp, or Ollama.

For example, with Ollama:

ollama run http://hf.co/ccore/Qwen3.5-4B-3D-Coder:Q4_K_M

One important setup detail: Disable Multi-Token Prediction (MTP) in LM Studio, as the developer reports that the default setting can cause generation problems.

📈 The model is still improving

The developer has been updating the released checkpoints throughout training.

The October 27 checkpoint reportedly delivers:

▪️ Higher-quality generated scenes

▪️ Better one-shot success rates

▪️ More coherent spatial composition

▪️ A fix for the earlier model layer issue

The author has also been comparing checkpoints using fixed prompts and seeds, evaluating both visual quality and whether the generated scenes actually run.

But there's an important distinction: these are preliminary developer evaluations, not independent coding benchmarks.

⚠️ It's still experimental

The model can produce impressive scenes, but it isn't consistently reliable.

▪️ Some generations contain JavaScript errors, including duplicate variable declarations.

▪️ Complex scenes may require repeated attempts or manual corrections.

▪️ It generates procedural Three.js geometry, not production-ready Blender meshes or CAD files.

▪️ Its general coding capabilities haven't yet been established through comprehensive public benchmarks.

▪️ The showcased examples don't guarantee the same quality for every prompt.

📦 What's next?

The developer plans to publish:

▪️ Full FP16 model weights

▪️ Additional quantized versions

▪️ The training dataset

▪️ A technical report comparing model checkpoints

At the time of this preview, the public repository contains the Q4_K_M and Q8_0 models.

💡 Why this matters

Small coding models have historically struggled with complex spatial reasoning and long, interconnected code generation.

CCORE is exploring whether specialized training can push a relatively tiny model much further.

And the implications go beyond 3D.

Imagine locally generating interactive website elements, game prototypes, procedural environments, educational simulations, or visual demonstrations without depending on a massive cloud model.

A 2.7 GB model file producing an entire interactive 3D scene is a pretty interesting direction for local AI.

🔗 Model, examples and downloads:
https://huggingface.co/ccore/Qwen3.5-4B-3D-Coder
#Qwen #LocalLLM #ThreeJS #AIcoding #OpenSourceAI