Creative Coding

A Solo Developer's 4-Billion-Parameter Model Turns a Photo Into a Working Three.js Scene

CCORE's Qwen3.5-4B-3D-Coder is an independent, Apache-2.0 fine-tune trained on 4,146 Three.js scenes with full reasoning traces — small enough to run locally, and now able to read a reference image and generate the matching 3D code.

Most of the AI-writes-3D-code conversation this year has centered on frontier labs — Google’s Three-Live, announced last week, verifies AI-generated Three.js inside a sandboxed VM before handing it off as portable AR. Qwen3.5-4B-3D-Coder, quietly built by a single developer going by CCORE (Caio Giovannetti), is the smaller, scrappier counterpart: an independent fine-tune of Alibaba’s Qwen3.5-4B, trained specifically to write Three.js, running at a size you can fit on a single consumer GPU.

Built on a hand-assembled dataset

The model’s model card is explicit that this is a personal experiment, not an official Qwen release: “an experiment exploring the feasibility of small language models for general-purpose coding and high-quality 3D scene generation.” It’s trained on ThreeJS-3D-Scenes, a companion dataset of 4,146 Three.js scene examples — individual objects, detailed interiors, architecture, landscapes, animated mechanisms, voxel environments — each one paired with a prompt, a full chain-of-thought reasoning trace, and the complete HTML/JavaScript that renders it. The dataset itself was generated with MiniMax’s M3.1 Flash model and released separately on Hugging Face under an MIT license, so the training data is as inspectable as the resulting model.

The system prompt baked into the model pins its output to a specific Three.js version (0.160.0) loaded via a single ES-module CDN URL, and explicitly bans addons or legacy examples paths — a small but telling detail that suggests real iteration against what actually breaks when a model free-associates import paths.

The part that’s new: show it a picture

The original release covered text-to-scene generation. The update that’s made it newsworthy this month is optional vision support: the Q4_K_M and Q8_0 GGUF quantizations can accept image inputs when paired with the original Qwen3.5-4B vision projector, so a user can attach a reference photo and ask for a Three.js scene built from it. In one published example, a reference photo of an orange voxel tree on a floating island, combined with the prompt “create this scene using voxels 3d,” produced a recognizable Three.js reconstruction — not a pixel-perfect copy, but a qualitatively faithful read of the reference’s composition and palette.

Small enough to actually run

The model ships as full FP16 safetensors plus Q4_K_M and Q8_0 GGUF quantizations, and CCORE’s own setup notes were tested locally with llama.cpp at a 65,536-token context window — well within reach of a single high-end consumer card, and a different proposition entirely from the multi-GPU infrastructure a frontier lab’s 3D-coding model assumes. At 368,000+ downloads and over 1,000 likes on Hugging Face as of this week, it’s also found a real audience among exactly the people most likely to want a local, inspectable Three.js assistant: solo creative coders who don’t want to route every sketch idea through someone else’s API.