Qwen3.5-4B-3D-Coder-MLX

Qwen3.5-4B-3D-Coder is an early-preview independent fine-tune of Qwen3.5-4B by ccore, exploring general-purpose coding and interactive 3D scene generation using self-contained Three.js HTML and JavaScript. It was trained on the MIT-licensed ccore Three.js 3D Scenes dataset, containing 4,146 examples with prompts, reasoning traces, complete code, and complexity labels. The repository includes FP16 safetensors, Q4_K_M and Q8_0 GGUF builds, and a technical report. An October 27 checkpoint reportedly improves one-shot generation and scene quality. No quantitative benchmark results are provided; evaluation is demonstrated through sample 3D scenes. Released under Apache-2.0, it includes recommended sampling settings and a Three.js-focused system prompt. LM Studio users should disable multi-token prediction. Some generations may still require retries or minor code corrections.

Repository Layout

prithivMLmods/Qwen3.5-4B-3D-Coder-MLX/
├── 8bit/
└── [root / bf16]

Use with mlx

Install the required library:

pip install -U mlx-vlm

BF16 Variant (Base Weights)

The unquantized BF16 mlx weights reside directly in the root repository path:

CLI (Terminal)

# 1. Text-to-3D Scene Generation
python -m mlx_vlm generate \
  --model prithivMLmods/Qwen3.5-4B-3D-Coder-MLX \
  --max-tokens 4096 \
  --temperature 0.6 \
  --prompt "Create a self-contained HTML file using Three.js that renders an interactive low-poly floating island with animated rotating windmills and OrbitControls."

# 2. Image-Conditioned 3D Scene Generation
python -m mlx_vlm generate \
  --model prithivMLmods/Qwen3.5-4B-3D-Coder-MLX \
  --max-tokens 4096 \
  --temperature 0.6 \
  --prompt "Recreate the 3D scene and lighting shown in this reference image as a complete, self-contained Three.js HTML file." \
  --image <path_to_image>

Python API

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_path = "prithivMLmods/Qwen3.5-4B-3D-Coder-MLX"
model, processor = load(model_path)
config = load_config(model_path)

prompt = "Create a self-contained HTML file using Three.js that renders an interactive low-poly floating island with animated rotating windmills and OrbitControls."
formatted_prompt = apply_chat_template(processor, config, prompt)

output = generate(
    model, 
    processor, 
    formatted_prompt, 
    max_tokens=4096, 
    temperature=0.6
)
print(output.text)

8-bit Variant

Access the 8-bit quantized files using --subfolder 8-bit:

CLI (Terminal)

# 1. Text-to-3D Scene Generation
python -m mlx_vlm generate \
  --model prithivMLmods/Qwen3.5-4B-3D-Coder-MLX \
  --subfolder 8-bit \
  --max-tokens 4096 \
  --temperature 0.6 \
  --prompt "Create a self-contained HTML file using Three.js that renders an interactive low-poly floating island with animated rotating windmills and OrbitControls."

# 2. Image-Conditioned 3D Scene Generation
python -m mlx_vlm generate \
  --model prithivMLmods/Qwen3.5-4B-3D-Coder-MLX \
  --subfolder 8-bit \
  --max-tokens 4096 \
  --temperature 0.6 \
  --prompt "Recreate the 3D scene and lighting shown in this reference image as a complete, self-contained Three.js HTML file." \
  --image <path_to_image>

Python API

from mlx_vlm import load, generate
from mlx_vlm.prompt_utils import apply_chat_template
from mlx_vlm.utils import load_config

model_path = "prithivMLmods/Qwen3.5-4B-3D-Coder-MLX"
model, processor = load(model_path, subfolder="8-bit")
config = load_config(model_path, subfolder="8-bit")

prompt = "Create a self-contained HTML file using Three.js that renders an interactive low-poly floating island with animated rotating windmills and OrbitControls."
formatted_prompt = apply_chat_template(processor, config, prompt)

output = generate(
    model, 
    processor, 
    formatted_prompt, 
    max_tokens=4096, 
    temperature=0.6
)
print(output.text)

License and Attribution

This model is based on and/or incorporates the following open-source projects and models:

• Qwen3.5-4B (Base): https://huggingface.co/Qwen/Qwen3.5-4B
• CCORE Qwen3.5-4B-3D-Coder: https://huggingface.co/ccore/Qwen3.5-4B-3D-Coder
• mlx-vlm: https://github.com/Blaizzy/mlx-vlm
• MLX: https://github.com/ml-explore/mlx

This model is released under the Apache License 2.0.