Qwen3.5-4B-GLB-3D-Coder: Text-to-GLB Game Assets on a 4B Apache Model

Qwen3.5-4B-GLB-3D-Coder is an Apache 2.0 fine-tune of Qwen/Qwen3.5-4B that targets a narrow but practical job: turn an English asset brief into a Three.js web page you can open locally and export as GLB for Godot or other editors. CCORE published the weights, GGUF builds, and training dataset on Hugging Face with a November 5, 2026 project release note.

This is a builder-focused release, not a frontier general model. The value is the loop: prompt, HTML scene, browser preview, GLB file. If you followed our October preview of Qwen3.5-4B-3D-Coder, treat this as the production drop with export, dataset, and gameplay showcase.

TL;DR

What CCORE actually shipped

The model card describes GLB-3D-Coder as specialized for low-poly game assets and procedural scenes. The workflow:

1. You send a text request describing a character, prop, building, or environment.
2. The model returns complete HTML and JavaScript using Three.js, including preview canvas and an Export GLB control.
3. You save the HTML, open it in a browser, click export, and import the GLB elsewhere.

The November 5 release adds pieces the October preview lacked:

The card warns anyone who downloaded before November 5, 2026 to delete old GGUF files and fetch again, because filenames did not change but weights did.

System prompt and generation settings

CCORE documents a fixed system instruction oriented at Godot users:

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.

Sampling defaults on the card:

The card stresses that training examples reached 72K tokens, averages sit around 20K, and 8K output limits break generations mid-thought or mid-HTML. That matters if you route the model through a conservative API gateway.

How to run it locally

GGUF path (fastest try)

1. Download Qwen3.5-4B-GLB-3D-Coder-Q4_K_M.gguf (about 2.71 GB) or the Q8_0 variant (about 4.48 GB) from the repo files tab.
2. Load in LM Studio, llama.cpp, or another Qwen3.5-compatible GGUF host.
3. Paste the system prompt from the card.
4. Send a short English prompt such as: Create a red barn with wooden fences and hay bales.
5. Save the assistant's HTML to disk and open it in Chrome or Firefox.
6. Click Export GLB and import into Godot.

The card notes the HTML may load Three.js from a CDN. Air-gapped runs need you to swap CDN URLs for local copies.

Safetensors path (fine-tuning)

Use the five-shard checkpoint with the repo config, tokenizer, and chat template if you plan to fine-tune or serve with transformers. Expect FP32 text weights plus BF16 vision tensors per metadata.

Agent harness path

The RIFTWARD Space includes a starter prompt for wiring a coding agent to a local model API. That matches how many readers already use Claude Code skills for 3D or isometric skill demos. GLB-3D-Coder is the model backend; your harness still owns file writes and browser tests.

3D world model cube illustrating Qwen3.5-4B GLB-3D-Coder text-to-asset workflow

Dataset: ThreeJS-3D-GLB-Assets

CCORE published ThreeJS-3D-GLB-Assets with 3,918 rendered examples across 40 domains, each tested for GLB export. That dataset is the main research artifact for reproducibility:

• You can inspect prompt and HTML pairs without trusting marketing screenshots alone.
• You can fine-tune rival small coders on the same corpus to test whether export quality is model-specific.
• You can filter domains (vehicles, bosses, laboratory props) to specialize a downstream LoRA.

Our earlier Scenario 3D model collection post covered a different lineage: static CC BY meshes, not executable HTML generators. The skills overlap in Godot import, not in training data format.

Showcases and comparisons on the card

CCORE documents three credibility layers, each with different evidential strength:

1. Side-by-side prompts versus DeepSeek V4.1 Flash, with CCORE on the right. Examples include a coral-armored lighthouse keeper, a frozen goblin camp, and long-form character briefs.
2. RIFTWARD, a browser arena built from generated assets plus orchestration code not produced by the 4B model alone.
3. Community demo from Naoki Kishida using an earlier preview checkpoint.

None of these replace a blind benchmark. They show what the author could reproduce on chosen prompts. For honest evaluation, pick ten held-out prompts from the dataset test split if CCORE publishes splits, or write your own.

Post-training polish concept for a small Qwen3.5-4B GLB-3D-Coder fine-tune

Hardware and runtime expectations

A 4B Q4_K_M file fits consumer laptops with 8 GB VRAM or unified memory, similar to other small local coders readers already run. The bottleneck is output length, not inference width. Long scenes spike RAM during generation.

If you compare against cloud APIs, price the token volume: twenty-thousand-token completions add up when you iterate art direction.

Limitations the card admits

CCORE lists frank limits:

• Some requests need retry or a small manual fix in the HTML.
• Materials, lighting, and animations may differ between browser preview and Godot import.
• Thinking-disabled or truncated runs are untested by the author.

That is healthier than silent failure. Plan a human review step before shipping generated assets in a commercial title, same as with any coding agent output.

Roadmap and funding note

The card asks for community support via Ko-fi and crypto wallets to fund a 100,000-example follow-on corpus and iterative tool-using trainers. That is aspirational; no delivery date is binding. Treat the plea as context on sustainability, not as a product guarantee.

How it compares to other 3D AI paths

For Three.js specifically, builders should still read official Three.js documentation when debugging exporters. The model speeds first draft, not license compliance or performance profiling.

Troubleshooting exports and bad HTML

When a generation fails, the failure mode is usually visible in the browser console, not in the model card.

CCORE recommends thinking mode on and warns against disabling it without testing. If your runtime strips chain-of-thought tokens, switch hosts or update templates before blaming the weights.

Security hygiene matters too. Generated HTML can pull scripts from CDNs. Review network calls before opening untrusted prompts on a machine with credentials logged in. For agent pipelines that auto-open files, put the step behind the same sandbox you use for Claude Code remote sessions.

When a 4B coder is the wrong tool

Skip GLB-3D-Coder when you need photoreal assets, CAD precision, or rigged characters with bone weights validated for film. It targets stylized low poly props and modular kit pieces. For narrative worlds with consistent art direction across dozens of assets, you still want a human art bible and probably a larger closed model for hero meshes.

It shines when you want iterative blocking: greybox a level in an afternoon, hand-polish five hero assets, and keep the rest procedural. That mirrors how the RIFTWARD Space on Hugging Face combined generated GLBs with orchestration code rather than claiming the 4B model shipped a full game alone.

What to try first

Copy-paste prompts from the card domains table if you are new:

• Create a village blacksmith wearing a blue apron.
• Create a stone dungeon doorway module with an iron frame.
• Create a wooden gate with an opening animation.

For a harder test, use a multi-paragraph brief naming materials, counts, and colors, mirroring the magnetic-rail engineer example in the showcase section. Success means a GLB that imports with sane scale; beauty is secondary.

Related reading

• October Qwen3.5-4B-3D-Coder preview
• Scenario's open 3D model collection
• Claude Code 3D simulator skill teardown
• Isometric Claude Code skill anatomy
• Tailsteer multi-turn image editing checkpoints
• LeRobot quickstart for embodied stacks
• Addy Osmani's checklist for agent output review

Model card details are accurate as of November 7, 2026, per the CCORE Hugging Face repository.