A prompt becomes a scene, not a mystery mesh
A new 4-billion parameter AI model from CCORE promises to streamline 3D asset creation for game developers. Unlike other AI tools that output opaque mesh files, this one generates standalone HTML and JavaScript scenes built with the Three.js library. The core idea is simple: you get code you can actually inspect and edit, not a black-box mesh.
Matthew Berman's video demonstrates the workflow. You type a natural-language request, like "Create a mana potion." The AI then writes a complete HTML file. Opening this file in any web browser reveals the 3D scene, which you can rotate and examine. A built-in button exports the asset as a GLB file.
This approach offers a significant practical payoff for indie developers. The generated GLB can be directly imported into game engines like
- Godot
- Unity
- Blender
Meanwhile, the underlying Three.js code remains fully accessible. This means you can tweak the asset's geometry, materials, or animations directly, offering a level of control absent in most other AI-generated 3D content. It’s about getting a usable asset and the blueprint.
Four billion parameters, built for the indie bottleneck
Built by CCORE, Qwen3.5-4B-GLB-3D-Coder is a specialized model based on Alibaba’s open-source Qwen3.5-4B. Released under the Apache 2.0 license, this 4-billion parameter AI is tuned specifically for 3D coding, not just raw mesh generation. It’s designed to tackle the common bottleneck indie game developers face when creating unique assets.
Crucially, the model emphasizes a local-first approach. Available in GGUF Q4_K_M and Q8_0 formats, alongside Safetensors, it runs on consumer-grade hardware. This means you can experiment and generate assets without relying on expensive cloud services, a significant win for budget-conscious creators.
Model’s training covers 25 distinct game-asset categories, offering a wide range of creative possibilities. It can generate anything from:
- Props
- Weapons
- Vehicles
- Modular environments
An optional vision projector even allows you to feed it image references, adding another layer of creative control. These features make it a powerful tool for quickly populating your game worlds with custom 3D content.
Why code can beat a prettier AI mesh
So why might you want code over a prettier AI-generated mesh? Text-to-3D diffusion tools often produce dense, noisy meshes that are difficult to edit. This model, by contrast, builds assets using procedural Three.js code, which means simpler, primitive-based geometry that’s lightweight and easier to modify.
Inspectable code and direct GLB export fit seamlessly into existing game development workflows, especially for indie studios. This makes it ideal for rapid asset blocking, prototyping, and creating modular game content that can be easily dropped into engines like Godot. Instead of a black-box mesh, you get human-readable JavaScript.
The goal isn't to build a finished game, but to accelerate asset creation. Early proof-of-concept examples, like the browser game RIFTWARD and various community RPG demos, show developers experimenting with the model's output. These showcases highlight its potential for generating functional game elements.
This approach prioritizes utility and editability over raw visual fidelity. For more technical details and to explore the model yourself, check out ccore/Qwen3.5-4B-GLB-3D-Coder on Hugging Face. It’s about fitting the tool to the task, not just generating impressive images.
Enjoying this? Get one like it in your inbox each morning.
one email a day · unsubscribe in two clicks · no third-party tracking
The catch: generated code still needs a human
Generated code, however, still needs a human touch. While recent checkpoints have improved reliability and geometry, the model can still produce syntax errors, broken GLB export logic, or missing texture bindings. Expect to debug generated Three.js for a while.
The tool also works best for specific styles. It excels at clean, low-poly rigid objects: think crates, weapons, or simple architecture. Don’t expect detailed anatomy, organic characters, or photorealistic sculpts; this AI isn't replacing a character artist.
Ultimately, try this for rapid prototyping and generating editable game props. It’s ideal for quickly populating a scene or brainstorming variations, especially for engines like Godot. But keep an artist or developer in the loop to refine the output.
CCORE’s commitment to open datasets and plans for a significantly larger training run—targeting 100,000 samples—suggests ongoing development. This is a tool that’s still growing, but it’s already a practical asset for indie developers.
Frequently Asked Questions
What is Qwen3.5-4B-GLB-3D-Coder?
It is an open-source 4-billion-parameter model that generates Three.js HTML and JavaScript scenes with GLB export.
Can it run locally?
Yes. GGUF versions are designed for local use through tools such as LM Studio or Ollama, including on some consumer laptops and GPUs.
Can I use its GLB assets in Godot?
Yes. The exported GLB files can be imported into Godot, as well as compatible tools such as Unity and Blender.
What kinds of assets does the model handle best?
It is strongest at stylized, low-poly props, weapons, modular architecture, and other rigid assets; complex organic characters are more challenging.
Does it generate finished 3D meshes directly?
It generates Three.js code that builds the scene procedurally, then exports the result as a GLB file.