Early Release: This is an initial experimental release of GEMMA-3D-CODER. Future updates will feature improved prompt adherence, expanded 3D asset generation formats, and better multi-turn context retention.

GEMMA-3D-CODER

GEMMA-3D-CODER is an experimental open-weight language model fine-tuned specifically for general-purpose coding and high-quality 3D scene / game asset generation using Three.js and web-standard technologies.

It is trained to understand natural language prompts and translate them into robust, self-contained HTML/JavaScript scenes or GLB-compatible game assets tailored for engines like Godot.

---

Example Prompts & Capabilities

This model excels at generating interactive, low-poly, voxel-style 3D scenes and procedural scripts via Three.js from CDN (cdnjs)[cite: 1].

Example 1: treasure chest

User: Create a treasure chest that can open and close

Output:

Example 2: Adventurer Utility Belt

User: Please create one self-contained HTML file with Three.js from cdnjs showing an Adventurer utility belt. Output:

GGUF DOWNLOAD

Format	Download	Notes
GGUF · IQ4_XS	GEMMA-3D-CODER.IQ4_XS.gguf	Extremely small 4-bit importance matrix quantization.
GGUF · Q2_K	GEMMA-3D-CODER.Q2_K.gguf	2-bit k-quant for maximum size reduction.
GGUF · Q3_K_L	GEMMA-3D-CODER.Q3_K_L.gguf	Large 3-bit k-quant variant.
GGUF · Q3_K_M	GEMMA-3D-CODER.Q3_K_M.gguf	Medium 3-bit k-quant variant.
GGUF · Q3_K_S	GEMMA-3D-CODER.Q3_K_S.gguf	Small 3-bit k-quant variant.
GGUF · Q4_K_M	GEMMA-3D-CODER.Q4_K_M.gguf	Medium 4-bit k-quant (recommended balance).
GGUF · Q4_K_S	GEMMA-3D-CODER.Q4_K_S.gguf	Small 4-bit k-quant variant.
GGUF · Q5_K_M	GEMMA-3D-CODER.Q5_K_M.gguf	Medium 5-bit k-quant for higher fidelity.
GGUF · Q5_K_S	GEMMA-3D-CODER.Q5_K_S.gguf	Small 5-bit k-quant variant.
GGUF · Q6_K	GEMMA-3D-CODER.Q6_K.gguf	6-bit k-quant for near-lossless performance.
GGUF · Q8_0	GEMMA-3D-CODER.Q8_0.gguf	8-bit quantization for high precision.
GGUF · f16	GEMMA-3D-CODER.f16.gguf	Unquantized 16-bit floating point weights.
Multimodal Projector · Q8_0	GEMMA-3D-CODER.mmproj-Q8_0.gguf	8-bit vision/multimodal projector weights.
Multimodal Projector · f16	GEMMA-3D-CODER.mmproj-f16.gguf	16-bit unquantized vision/multimodal projector weights.

🛠️ Quick Start (Usage with Transformers)

You can load and use this model directly using Hugging Face transformers:

from transformers import AutoProcessor, AutoModelForCausalLM
import torch

processor = AutoProcessor.from_pretrained("Shrijanagain/GEMMA-3D-CODER", trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    "Shrijanagain/GEMMA-3D-CODER",
    torch_dtype=torch.float16,
    device_map="auto"
)

messages = [
    {
        "role": "system",
        "content": "Create game assets for the Godot editor, which uses GLB files. Create whatever the user requests as a good-quality assets."
    },
    {
        "role": "user",
        "content": "Create a low-poly goblin game asset using Three.js core module."
    }
]

prompt = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

inputs = processor(text=prompt, return_tensors="pt").to("cuda")

outputs = model.generate(
    **inputs,
    max_new_tokens=4096,
    use_cache=True,
    temperature=0.7
)

print(processor.batch_decode(outputs, skip_special_tokens=True)[0])


**System prompt for Three.js scenes**

Create high-quality, production-ready low-poly 3D game assets specifically optimized for the Godot engine (which uses the GLB/GLTF format).

Requirements:

• Build the asset using only Three.js (latest stable version via CDN).
• Deliver everything as a single, complete, standalone HTML file. No external dependencies except the Three.js CDN, no build steps, no additional files.
• The HTML page must:
  • Immediately display the 3D asset in a clean, well-lit scene with proper camera controls (OrbitControls).
  • Include a clear, visible “Export GLB” button.
  • When the button is clicked, generate and download a fully working .glb file that can be directly imported into Godot without errors.
• Asset style: Clean, good-quality low-poly aesthetic suitable for games (clean topology, reasonable poly count, proper UVs if needed, consistent scale, correct origin/pivot).
• If the requested asset logically needs animation (e.g., character idle, rotating propeller, opening door, bouncing object, etc.), implement a smooth looping animation and make sure the animation is correctly exported inside the GLB (including animation tracks).
• If no animation is needed, export a clean static GLB.
• Use proper materials, lighting, and basic PBR where appropriate so the asset looks good both in the Three.js preview and after import into Godot.
• Keep the code well-structured, readable, and commented so it can be easily modified later.
• Handle the GLB export correctly using Three.js’s GLTFExporter (including animations when present).

When the user describes an asset (or simply says what they want), generate the complete HTML file that fulfills all the above requirements.

### Output length and thinking

| Parameter | Recommended setting |
|:---|:---|
| Maximum output tokens | **Unlimited — `-1` where supported** |
| Thinking / reasoning | **Enabled, with no effort limit or separate thinking-token budget** |

**An output limit is too restrictive for these generations** and can stop the response during thinking or before the HTML is complete, leaving incomplete or unusable code. Set the output limit to **unlimited (`-1`)** in runtimes that support it.

Unlimited output removes the explicit generation cap; the runtime's context and resource limits still apply.