CCORE has released GLB-3D-Coder, a 4-billion-parameter language model that turns a text description into an HTML and JavaScript page that generates a three-dimensional object. The page uses Three.js, a library for 3D graphics in the browser, shows a live preview and lets users export the result as a GLB file, the binary form of glTF supported by Godot, Blender and other 3D tools.
The difference from the more common 3D generators lies in how the geometry is built. The model writes the code that creates the shape, so every asset can be inspected, edited line by line and tracked with version-control systems such as Git. Developers can change dimensions, materials, colours, lights and scene layout directly in the generated code.
The model is a full fine-tune of Qwen3.5-4B, Alibaba's open architecture, and is distributed under the Apache 2.0 licence. Full Safetensors weights are available, along with two quantised GGUF versions, Q4_K_M and Q8_0. The Q4_K_M version takes up about 2.7 GB and can run locally with llama.cpp, Ollama, LM Studio or vLLM. According to its Hugging Face page, the model recorded about 20,000 downloads in its first week.
The model was trained on 3,918 rendered examples, validated as GLB files, spanning 40 domains. That dataset was released alongside the model. CCORE also points to two larger collections, of 4,146 and 3,445 Three.js scenes, published on Hugging Face; these figures come from the developer alone.
The release includes a demo, RIFTWARD, a browser arena in which every asset was generated by the model. According to the release notes, the model supports animation: when a prompt calls for movement, the generated page can define animation clips and include them in the exported GLB. Some community users have noted that outputs may remain limited to static assets or need manual correction for complex tasks, such as assets requiring specific skeletal rigs.
A practical limit concerns response length, which averages around 20,000 tokens. An 8,000-token output cap truncates generation and breaks the HTML. The project's stated next milestone is a 100,000-example dataset for small models that can test and refine their own outputs.
For those working on games and web environments, the release offers a starting point: a small model that runs on home hardware and produces assets in a standard, editable format. CCORE announced the release on Reddit, where the post gathered about 130 upvotes.

No comments yet. Be the first!