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.
The Q4_K_M and Q8_0 GGUF versions can also accept image inputs when loaded with the original Qwen3.5-4B vision encoder/projector. You can attach a reference image and ask the model to generate a Three.js scene from it.
The reference image was supplied in community discussion #9; credit for the reference belongs to its original creator. The right-hand image is a screenshot of the generated scene. This is a qualitative example of image-guided generation; the output interprets the reference and does not reproduce every detail.
Download a model GGUF and the projector, then load both files explicitly:
llama-server \
-m Qwen3.5-4B-3D-Coder-Q4_K_M.gguf \
--mmproj mmproj-Qwen3.5-4B-BF16.gguf \
--n-gpu-layers all --flash-attn on \
--ctx-size 65536 --parallel 1
Open the server's WebUI, attach an image, and send your prompt. The setup was tested locally with llama.cpp b11889. For Q8_0, change the model filename to Qwen3.5-4B-3D-Coder-Q8_0.gguf; --cache-type-k q8_0 --cache-type-v q8_0 can reduce KV-cache memory use and was tested locally with vision at a configured 64K context. These were short tests, not full-context stress tests.
The projector uses additional memory when loaded, and image tokens share the context window with the prompt and generated code. Image support in other frontends depends on their multimodal implementation.
Projector provenance: this is the unmodified BF16 projector from lmstudio-community/Qwen3.5-4B-GGUF, based on Qwen/Qwen3.5-4B, released under Apache 2.0. It is an optional companion file for the existing GGUF weights.
The original FP16 weights are now available in this repository in Safetensors format, alongside the Q4_K_M and Q8_0 GGUF versions.
The release includes all four weight shards, the model configuration, tokenizer, generation configuration, chat template, and image/video processor settings. The FP16 weights were verified locally for loading and text generation with Transformers 5.20.2.
The FP16 generation defaults use temperature 0.6, top-k 20, top-p 0.95, and repetition penalty 1.0, with caching enabled.
It contains 4,146 scene examples (3,933 training / 213 validation), including prompts, full thinking traces, complete HTML/JavaScript responses, and complexity levels. Rendered reference images are available in the optional previews configuration. The dataset is released under the MIT License.
Both Q4_K_M and Q8_0 now contain the updated October 27 checkpoint. If you downloaded an earlier version, please download the files again to get this update.
In my latest tests, one-shot success has improved substantially, along with overall scene quality. Some scenes can still require another attempt or a correction.
Generation parameters remain unchanged for both quantizations. Use the settings listed below. The earlier fix for the layer-related issue reported by some users is also included.
The comparison below shows the Rome scene across the October 25, 26, and 27 previews.
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.
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.
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.
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 clearance to guide
the layout around the structures.
Render it as a deliberate 3D voxel model built from visible cubic blocks,
with clear depth, readable silhouettes, and solid spatial relationships.
Build a self-contained HTML scene with Three.js from cdnjs featuring a
dragonrider aerie on a sea stack, with perches, a hatchery, and a windlass.
The aerie occupies a flat-topped sea stack, reached by a rope windlass from
the shore. Four projecting stone perches ring the central hatchery, a
vaulted cave with sand nests and a smoke hole. A timber landing deck holds
a hoist crane and harness racks, while switchback stairs connect the
perches to a water-level boathouse.
Render it as a deliberate 3D voxel model built from visible cubic blocks,
with clear depth, readable silhouettes, and solid spatial relationships.
Make a standalone HTML scene with Three.js from cdnjs depicting an animated
air traffic control tower with rotating radar and light signals.
A tapered control tower rises from a rectangular base to a cantilevered
glass cab with angled glazing panels and an external catwalk. A primary
surveillance radar on a roof pedestal turns continuously, and a rotating
beacon sits above the cab.
I’ve also released an early preview of GLB-3D-Coder, focused on creating reusable 3D assets with GLB export for game editors such as Godot. I’m preparing its training dataset for public release, with roughly the same number of examples as the dataset released with this model.
Beyond that, I’m planning a larger dataset of 100,000 high-quality examples, with context lengths ranging from 8K to 128K tokens, targeting a training run roughly 15× larger than the current GLB model’s run. The first goal is reliable one-shot asset generation, followed by a version designed to work directly inside an agent harness, generating, testing, and refining assets.
As with this release, I plan to share the datasets and model weights openly. Community support would help fund the data generation, validation, and GPU compute needed to make these next steps possible.
I'm independently developing and releasing specialized AI models and datasets for 3D generation, procedural environments, and creative coding, with the goal of making these tools accessible to everyone.
Training new models, generating high-quality datasets, validating thousands of examples, and running experiments require significant GPU resources, time, and effort.
If you enjoy this project and would like to help me train more models, expand the datasets, and explore new capabilities, any contribution is greatly appreciated!
My goal is to keep sharing models, weights, datasets, and improvements openly with the community.
BAAI