Instructions to use ccore/Qwen3.5-4B-GLB-3D-Coder with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use ccore/Qwen3.5-4B-GLB-3D-Coder with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M
# Run inference directly in the terminal:
llama cli -hf ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M

Build from source code

git clone https://github.com/ggml-org/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-app
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama serve -hf ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M
# Run inference directly in the terminal:
./build/bin/llama cli -hf ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M

Use Docker

docker model run hf.co/ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M

Llama new
LM Studio
Jan

vLLM

How to use ccore/Qwen3.5-4B-GLB-3D-Coder with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ccore/Qwen3.5-4B-GLB-3D-Coder"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ccore/Qwen3.5-4B-GLB-3D-Coder",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M

Ollama
How to use ccore/Qwen3.5-4B-GLB-3D-Coder with Ollama:
```
ollama run hf.co/ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M
```
Unsloth Desktop

How to use ccore/Qwen3.5-4B-GLB-3D-Coder with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use ccore/Qwen3.5-4B-GLB-3D-Coder with Docker Model Runner:
```
docker model run hf.co/ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M
```

Lemonade

How to use ccore/Qwen3.5-4B-GLB-3D-Coder with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M

Run and chat with the model

lemonade run user.Qwen3.5-4B-GLB-3D-Coder-Q4_K_M

List all available models

lemonade list

Hermes Agent

How to use ccore/Qwen3.5-4B-GLB-3D-Coder with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use ccore/Qwen3.5-4B-GLB-3D-Coder with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "ccore/Qwen3.5-4B-GLB-3D-Coder:Q4_K_M" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

CCORE GLB-3D-Coder

Turn a text prompt into an editable 3D asset.

CCORE GLB-3D-Coder is a specialized 4B coding model for low-poly game assets and procedural scenes. Describe a character, prop, building, or environment, and the model writes a complete Three.js HTML/JavaScript page with a preview and an Export GLB button. Open the page, inspect the result, and export the asset for Godot or another editor that supports GLB.

CCORE project release · November 5, 2026 · Apache 2.0
Repository: ccore/Qwen3.5-4B-GLB-3D-Coder · Checkpoint: latest uploaded

This independently developed CCORE model was trained using Qwen/Qwen3.5-4B as its base. CCORE maintains this release.

Q4_K_M, Q8_0, and the original Safetensors weights are available now. The training dataset is planned for release in the next few days.

Important: If you downloaded the model before November 5, 2026, delete the old model files and download them again in LM Studio or whichever app you use. The GGUF filenames haven't changed, so make sure you're loading the updated weights.

Get the weights · Use the model · Prompt ideas · Upcoming releases

RIFTWARD — playable arena showcase

RIFTWARD: Guardians of the Ancient is a browser arena built with character, terrain, architecture and environmental assets generated with CCORE GLB-3D-Coder 4B using the latest uploaded checkpoint. The assets were assembled, refined and integrated into a playable game with assistance from an orchestration and validation harness, which also handled gameplay systems, interface, effects and testing.

Play RIFTWARD on Hugging Face Spaces · Starter prompt for any coding agent

The game runs entirely in the browser using exported GLB assets; no live model inference is needed to play. The Space includes a copyable prompt explaining how to connect a coding agent to a local model API and build a game from its outputs.

Community showcase

Naoki Kishida (@kis) created a playable 3D RPG demo using assets generated with the earlier preview version of GLB-3D-Coder.

Play Naoki Kishida's 3D RPG demo

Thanks to Naoki for sharing this project with the community!

Examples beyond the training data

Here are some examples the model created from prompts that were not included in its training dataset, ranging from short requests to detailed asset descriptions. The images below come from the project's showcase video.

Comparison setup: Both models received the same prompts, with DeepSeek set to high reasoning effort. In each comparison, DeepSeek V4.1 Flash 1017 is on the left and CCORE GLB-3D-Coder (4B) is on the right.

Coral-armored lighthouse keeper

Prompt for both models: Create a coral-armored lighthouse keeper.

Frozen goblin watch camp

Prompt for both models: Create a frozen goblin watch camp.

Magnetic-rail rescue engineer

Prompt used for both models

Create a magnetic-rail rescue engineer character. Show an adult specialist in a grounded stance wearing a charcoal work suit, orange safety panels, insulated gloves, a compact helmet, and heavy boots with visible magnetic sole plates. Equip the character with a coil-testing wand, a folded emergency rail clamp, a shoulder radio, and a silver power pack connected to the tool belt by a thick blue cable. Make the proportions, cable route, pack mounting, tools, and protective equipment mechanically plausible. Use a plain presentation without extra scenery. Keep it static.

Bioluminescent cave tram station

Prompt used for both models

Create a compact bioluminescent cave tram station scene. Include a short raised platform with safety rails, a small two-seat mine tram on narrow rails, a hand-operated cable winch, a ticket pedestal, a bench, and six blue glowing crystal lamps mounted in protective cages. Use dark stone, weathered timber, black iron rails, teal tram panels, and cyan crystals. Make the wheel flanges, rail spacing, cable route, platform supports, lamp cages, and passenger access mechanically coherent. Use a plain presentation without surrounding cave walls or extra scenery. Keep everything static.

Other examples


Animated laboratory microscope · Short prompt Prompt: `Create a white laboratory microscope with a rotating objective turret.`	Farmyard · Short prompt Prompt: `Create a small farmyard with a red barn, wooden fences, hay bales, a water trough, and a vegetable patch.`

The microscope's GLB export was adjusted to preserve its existing turret animation.

Weights and release status

The released Q4_K_M, Q8_0, and original Safetensors weights use the latest uploaded checkpoint.

Format	Download	Notes
GGUF · Q4_K_M	Qwen3.5-4B-GLB-3D-Coder-Q4_K_M.gguf	Compact quantized weights.
GGUF · Q8_0	Qwen3.5-4B-GLB-3D-Coder-Q8_0.gguf	Quantization used for the recorded evaluation.
Original Safetensors	Checkpoint files	Five shards, approximately 17.49 GB. FP32 text weights + BF16 vision weights. Available now — original, unconverted checkpoint.
Optional BF16 vision projector	mmproj-Qwen3.5-4B-BF16.gguf	644 MiB; needed only for image inputs in a compatible multimodal GGUF runtime.

The optional projector is the original, unmodified Qwen3.5-4B vision encoder/projector from lmstudio-community/Qwen3.5-4B-GGUF, provided under Apache 2.0.

Quick start

Load a GGUF in a runtime that supports Qwen3.5, or load the original checkpoint with its configuration, tokenizer, and chat template.
Set the system prompt and sampling parameters below.
Send an English asset request. Start with a short idea, or specify the parts, materials, colors, layout, and animation you need.
Save the complete generated HTML as an .html file and open it in a browser. The page may load Three.js and its add-ons from a CDN.
Inspect the scene, click Export GLB, and import the downloaded .glb into your editor.

Request animation explicitly when you need it, and check that it is included in the exported GLB.

Sampling settings

Parameter	Value
Temperature	0.8
Top K	40
Top P	0.95
Min P	0.0
Repeat penalty	1.0

The recorded evaluation used these settings with thinking enabled. Allow enough output tokens for the complete HTML; detailed scenes can produce long responses.

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

Allow the model to finish generating. The model was trained on examples reaching 72K tokens. Responses average around 20K tokens, and some can be longer.

An 8K 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.

Keep thinking enabled, without limiting reasoning effort or imposing a separate thinking-token budget. I have not validated generation quality with thinking disabled, truncated, or restricted.

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

System prompt

Create game assets for the Godot editor, which uses GLB files. Create whatever the user requests as a good-quality low-poly asset using only Three.js in one complete standalone HTML file. The HTML must show the asset and include an Export GLB button that downloads a working GLB. If the requested asset needs animation, add it and include the animation in the exported GLB; otherwise make it static.

Domains and prompt ideas

Use these categories to explore the model and shape your requests.

Domain	Prompt idea
Characters and NPCs	`Create a village blacksmith wearing a blue apron.`
Enemies	`Create a swamp goblin scout with a wooden shield.`
Bosses	`Create a stone golem boss with glowing green crystals.`
Companions and creatures	`Create a tiny fox companion wearing a red scarf.`
Weapons and armor	`Create a short iron sword with a leather-wrapped grip.`
Items, tools, and pickups	`Create a blue mana potion in a hexagonal glass bottle.`
Props and furniture	`Create a weathered wooden workbench with a small lantern.`
Nature and environment	`Create a pine tree growing on a rocky mound.`
Modular architecture	`Create a stone dungeon doorway module with an iron frame.`
Interactive objects	`Create a wooden gate with an opening animation.`
Ground transport	`Create a compact courier motorcycle with cargo boxes.`
Air and space transport	`Create a small cargo airship with twin propellers.`
Watercraft	`Create a weathered fishing skiff with nets and wooden crates.`
Industrial machinery	`Create a compact mining conveyor with metal support legs.`
Laboratory and medical assets	`Create a portable botanical analysis station with sample jars.`
Urban services	`Create a modern bus stop with a bench and route sign.`
Sports and playgrounds	`Create a small playground climbing wall with colorful holds.`
Farming and homestead	`Create an orchard harvest cart filled with wooden fruit crates.`
Cuisine and markets	`Create a bakery display counter with bread baskets.`
Museums and field research	`Create an ancient artifact display case on a stone pedestal.`
Traps and hazards	`Create a hidden spiked floor trap with stone tiles.`
Festivals and ceremonies	`Create a harvest festival booth with colorful banners.`
Battlefield scenery	`Create a reinforced frontier command post with sandbag defenses.`
Underwater structures	`Create a coral observation dome with a small docking platform.`
Puzzles and gameplay markers	`Create a glowing rune sequence console with three stone buttons.`

For a detailed request, name the main object first, then describe its components, colors, materials, placement, and intended movement. The exact prompts in the gallery provide examples of longer briefs.

Current limitations

Some requests need a retry or a small code correction. Inspect the exported asset in your editor, since materials, lighting, and animations can differ from the browser preview.

What comes next

The training dataset will follow in the next few days. Beyond this release, I want to build larger, carefully reviewed datasets and more capable small models, including specialized models in other areas.

I've developed and funded this work myself, and I'm releasing the models, weights, and datasets free of charge to the community.

Next, I'm planning 100,000 high-quality examples for 3D and general-purpose coding harnesses to train new small models that can use tools, test their outputs, and refine their work through iteration.

I expect creating these datasets and training the new models to require around 20 times the effort of the current project. I need the community's support to make this next stage possible.

Support the project

I cover the cost of data preparation, validation, experiments, and GPU training myself. If this model is useful to you and you want to help fund larger datasets and future open releases, you can support the work directly: