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

CCORE GLB-3D-Coder (ccore/Qwen3.5-4B-GLB-3D-Coder, checkpoint u008186) is an open-weight (Apache-2.0), specialized 4B-parameter coding model fine-tuned from Qwen/Qwen3.5-4B that turns English text prompts into editable, low-poly 3D game assets and procedural scenes. Given a short or detailed description of a character, prop, vehicle, building, or environment, the model generates a complete standalone Three.js HTML/JavaScript page featuring a live browser preview, optional animations, and an Export GLB button to download ready-to-use .glb files for game engines like Godot. Evaluated with thinking enabled (temperature=0.8, top_k=40, top_p=0.95, repeat_penalty=1.0) and showcased against DeepSeek V4.1 Flash 1017 on unseen complex prompts, it is available in compact Q4_K_M and Q8_0 GGUF quantizations, original 5-shard Safetensors (~17.49 GB with FP32 text and BF16 vision weights), and an optional 644 MiB BF16 mmproj vision projector for multimodal GGUF runtimes.

Repository Layout

prithivMLmods/Qwen3.5-4B-GLB-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-GLB 3D Game Asset Generation
python -m mlx_vlm generate \
  --model prithivMLmods/Qwen3.5-4B-GLB-3D-Coder-MLX \
  --max-tokens 4096 \
  --temperature 0.8 \
  --prompt "Create a low-poly stone golem boss with glowing cyan crystals and an Export GLB button in a standalone Three.js HTML file."

# 2. Image-Conditioned 3D Asset Generation
python -m mlx_vlm generate \
  --model prithivMLmods/Qwen3.5-4B-GLB-3D-Coder-MLX \
  --max-tokens 4096 \
  --temperature 0.8 \
  --prompt "Build a low-poly 3D game asset matching this reference image in Three.js with an Export GLB button." \
  --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-GLB-3D-Coder-MLX"
model, processor = load(model_path)
config = load_config(model_path)

prompt = "Create a low-poly stone golem boss with glowing cyan crystals and an Export GLB button in a standalone Three.js HTML file."
formatted_prompt = apply_chat_template(processor, config, prompt)

output = generate(
    model, 
    processor, 
    formatted_prompt, 
    max_tokens=4096, 
    temperature=0.8
)
print(output.text)

8-bit Variant

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

CLI (Terminal)

# 1. Text-to-GLB 3D Game Asset Generation
python -m mlx_vlm generate \
  --model prithivMLmods/Qwen3.5-4B-GLB-3D-Coder-MLX \
  --subfolder 8-bit \
  --max-tokens 4096 \
  --temperature 0.8 \
  --prompt "Create a low-poly stone golem boss with glowing cyan crystals and an Export GLB button in a standalone Three.js HTML file."

# 2. Image-Conditioned 3D Asset Generation
python -m mlx_vlm generate \
  --model prithivMLmods/Qwen3.5-4B-GLB-3D-Coder-MLX \
  --subfolder 8-bit \
  --max-tokens 4096 \
  --temperature 0.8 \
  --prompt "Build a low-poly 3D game asset matching this reference image in Three.js with an Export GLB button." \
  --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-GLB-3D-Coder-MLX"
model, processor = load(model_path, subfolder="8-bit")
config = load_config(model_path, subfolder="8-bit")

prompt = "Create a low-poly stone golem boss with glowing cyan crystals and an Export GLB button in a standalone Three.js HTML file."
formatted_prompt = apply_chat_template(processor, config, prompt)

output = generate(
    model, 
    processor, 
    formatted_prompt, 
    max_tokens=4096, 
    temperature=0.8
)
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-GLB-3D-Coder: https://huggingface.co/ccore/Qwen3.5-4B-GLB-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.