What happened

Hugging Face's trending models list now features ccore/Qwen3.5-4B-3D-Coder at number 26 (166 likes, 3,544 recent downloads). It is a 4B-parameter model fine-tuned by an independent developer, CCORE, on Qwen/Qwen3.5-4B, released as an "early preview." The author explicitly states it is not an official Qwen release, but a personal experiment shared mid-training because initial hands-on results looked promising.

Key details

• Goal: explore the feasibility of small language models for general-purpose coding and high-quality 3D scene generation, with a particular focus on interactive Three.js scenes.
• Update: on October 26, 2026, both Q4_K_M and Q8_0 were updated to fix a layer-related issue reported by users; per the author's tests, the update appears to produce more successful one-shot scene generations.
• Schedule: training is expected to finish on October 28, 2026, at 07:30 UTC; the Q8_0 GGUF will be updated to the latest checkpoint on October 26 and 27. After training, the author plans to release full FP16 weights, all quantized versions, the training dataset, and an evaluation report covering a variety of prompts.
• Usage notes: recommended settings are Temperature 0.6–0.8, Top P 0.90–0.95, Top K 20–40, repeat penalty 1.0. In LM Studio, multi-token prediction (MTP) must be disabled for the model to work properly, since LM Studio enables it by default.
• Scene generation: the README provides a system prompt restricting the model to the Three.js 0.160.0 core ES module from cdnjs, with no addons or legacy examples/js imports and camera controls implemented directly in the HTML.

Who it affects

• Frontend and creative-coding developers who want to generate interactive Three.js 3D scenes with a local small model can try the GGUF builds directly.
• The open-source community: the promised release of the training dataset and evaluation report is relevant to developers who want to reproduce the work or fine-tune other models.
• Practitioners tracking small-model capability limits: the project is essentially a public experiment on whether a 4B model can handle general coding plus high-quality 3D scene generation.

Why it matters

The model matters on two fronts. First, its direction: it pushes a small (4B) model toward a concrete, verifiable task combining general coding and 3D scene generation, with README examples spanning stylized architecture, gardens, voxel environments, furniture, workshop interiors, and mechanical objects, all built through HTML and JavaScript as a visual reference of capability. Second, its openness: the early-preview approach with rolling checkpoint updates, followed by the dataset and evaluation report, offers the community a transparent look at how such a fine-tune is built. That said, it remains a preview of a model still in training; the Q4_K_M and Q8_0 files may continue to change, and a fuller assessment should wait for the post-training evaluation report.

Hugging Face 热门模型第 26 名：ccore/Qwen3.5-4B-3D-Coder，166 赞，近期下载 3544 次。 标签：gguf, qwen3_5, code, threejs, 3d, preview, base_model:Qwen/Qwen3.5-4B, base_model:finetune:Qwen/Qwen3.5-4B, license:apache-2.0, endpoints_compatible, region:us, conversational

README：

Qwen3.5-4B-3D-Coder

Early preview · General coding with a focus on 3D

An experiment exploring the feasibility of small language models for general-purpose coding and high-quality 3D scene generation.

This is an early preview of a 4B model I am still training, with a particular focus on creating interactive 3D scenes with Three.js. Initial hands-on results have been promising for coding in general, and especially for generating 3D scenes, so I decided to share the progress while development continues.

This is an independent fine-tune by CCORE, based on Qwen/Qwen3.5-4B. It is not an official Qwen release.

Update — October 26, 2026

Both Q4_K_M and Q8_0 have been updated. If you downloaded an earlier version, please download the files again to get this update.

• Fixed the layer-related issue reported by some users.
• In my latest tests, this update appears to produce more successful one-shot scene generations.

The image below shows one example comparison from my tests.

Training and release schedule

Training is expected to finish on October 28, 2026, at 07:30 UTC.

On October 26 and 27, 2026, I will update the Q8_0 GGUF to the latest available training checkpoint.

After training finishes, I will publish:

• The full FP16 model weights.
• The full set of quantized versions.
• The training dataset, so others can use it to train or fine-tune other models.
• An evaluation report with a variety of prompts, showing how the model performed on them throughout training.

Early examples

Selected screenshots of rendered Three.js scenes: stylized architecture, gardens, voxel environments, furniture, workshop interiors, and mechanical objects, built through HTML and JavaScript.

Click any image to view the original screenshot at full size.

Sakura Temple Garden Sunset Village

Dragonrider Aerie Voxel launch site

The Bookbinding Bench LGA 3647 Socket

Pelican Bike No. 07 Bow-Front Secretary Desk

Kōrin Garden Sakura bonsai

Generation settings

Use the following settings for both the Q4_K_M and Q8 versions:

Parameter	Value
Temperature	0.6–0.8
Top P	0.90–0.95
Repeat penalty	1.0
Top K	20–40

LM Studio: disable MTP

Important: If you're running this model in LM Studio, disable multi-token prediction (MTP) for the model to work properly. LM Studio enables this option by default.

System prompt for Three.js scenes

Use only the Three.js core ES module from
https://cdnjs.cloudflare.com/ajax/libs/three.js/0.160.0/three.module.min.js.
Do not import addons or legacy examples/js paths;
implement any camera controls directly inside the HTML.

Example prompts

Use the system prompt above, then try one of these user prompts.

Elephant rehabilitation habitat

Please create one self-contained HTML file with Three.js from cdnjs showing
an elephant rehabilitation habitat.

A broad sheltered yard connects water, shade, and feeding areas through
accessible paths. Use the elephant's scale and movement cl

The original belongs to Hugging Face 热门模型 and its authors and is reproduced here for reading. To have it removed, write to support@aisa.one.