Independent author CCORE's preview programming model, fine-tuned from Qwen3.5-4B, generates self-contained HTML files with interactive Three.js 3D scenes based on text descriptions. It was first listed on 10 29. This re-listing is because, starting 10 30, the GGUF version supports reference images: give it an image, and it writes a corresponding 3D scene.
Rising score 348 · Likes 394 · Downloads 29615
Reasons
- New content 1: On 10 30, the author uploaded a visual projector file of approximately 644 MiB. It is the unmodified BF16 mmproj from the original Qwen3.5-4B, sourced from lmstudio-community. The Q4_K_M and Q8_0 quantized versions can process images once this file is loaded. The example in the card provides a reference image of a voxel tree island with the prompt "create this scene using voxels 3d", and the model generates the corresponding Three.js scene.
- New content 2: On 10 29, following the previous listing, the author published a training technical report PDF. The FP16 safetensors weights and the training dataset of 4,146 scenes (MIT license) are now also public.
- With a size of 4B, it can run locally using GGUF, under the Apache-2.0 license.
Usage
- llama.cpp: Run
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, and attach images with prompts in the built-in WebUI. The author tested it with llama.cpp b11889; for Q8_0, you can add--cache-type-k q8\_0 --cache-type-v q8\_0to save KV cache. - Use the system prompt as specified in the card: use only three.module.min.js from three.js 0.160.0 on cdnjs, do not import addons, and implement camera controls directly in HTML. For sampling, use temperature 0.6–0.8, top_p 0.90–0.95, and top_k 20–40.
- In LM Studio, disable Multi-Token Prediction (MTP), which is enabled by default.
Risks
- The image understanding capability comes from the projector of the original Qwen3.5-4B and has not been specifically trained for 3D scenes. The author states that the output reflects an understanding of the reference image rather than a pixel-perfect reconstruction; there is only one example, which underwent short testing without full-context stress tests.
- It remains an early preview with no benchmark scores. Image tokens and generated code share the context, and the projector requires additional memory; whether other frontends can process images depends on their respective multimodal implementations.
- This is a personal fine-tune, not an official Qwen release.