Qwen3.5-4B-3D-Coder: A 4B Model That Writes Three.js Scenes

On 28 October 2026 the independent developer CCore published full FP16 weights for Qwen3.5-4B-3D-Coder. It is an Apache 2.0 fine-tune of Qwen3.5-4B, explicitly not an official Qwen release, built for one job: turn a scene description into a self-contained HTML plus Three.js walkthrough. The same repository ships a 4,146-example training dataset

Tool Interface

Interactive tool will be available soon

Features

✓ 4B parameters under Apache 2.0, small enough for a local GPU or quantised CPU inference without queueing behind an API
✓ Output is the deliverable: a self-contained HTML file pulling Three.js from cdnjs, open it in a browser and walk around
✓ Open training data: the CCORE Three.js 3D Scenes dataset has 4,146 examples (3,933 train / 213 validation) with prompts, full thinking traces and HTML/JS answers under MIT
✓ Three weight formats: FP16 Safetensors across four shards, plus Q4_K_M at 2.71GB and Q8_0 at 4.48GB GGUF builds
✓ Sample scenes run from a sakura temple garden and sunset village to a voxel launch site and an LGA 3647 socket, so it expects detail-dense industrial prompts

How to Use

Tight on VRAM? Grab Q4_K_M and load it in LM Studio or llama.cpp, and disable Multi-Token Prediction as the card advises
Use the documented defaults: temperature 0.6, top-k 20, top-p 0.95 and repetition penalty 1.0
Write prompts down to the part level: materials, lighting, moving parts, camera path, because that is how the training data reads
Expect retries: the author says some scenes still need another attempt or a correction, so do not treat one-shot success as an SLA

FAQ

Is this an official Qwen release from Alibaba?

No. The model card states it is an independent fine-tune by CCORE based on Qwen/Qwen3.5-4B and not an official Qwen release. Source: https://huggingface.co/ccore/Qwen3.5-4B-3D-Coder

What licence applies, and can I use it commercially?

The model repository is tagged Apache 2.0 and the CCORE Three.js 3D Scenes dataset is MIT licensed. Re-check the current model and dataset cards yourself before commercial use.

How much hardware does it need?

Pick your weight format: Q4_K_M is 2.71GB, Q8_0 is 4.48GB, and FP16 spans four Safetensors shards. The author verified the FP16 weights load and generate text with Transformers 5.20.2.

How good are the scenes?

The author reports that one-shot success improved substantially with the 27 October checkpoint and that overall scene quality improved, while some scenes still need another attempt or a fix. That is self-reported, with no third-party benchmark.

Can I fine-tune it on my own style?

Yes. The dataset is public under MIT, and the author says publishing it was meant to let others train or fine-tune other models.