GLB-3D-Coder
A 4B-parameter code generation model trained on Qwen3.5-4B that generates Three.js scene code with GLB export, used to produce 3D assets directly importable into game engines from natural language descriptions.
Timeline (last 90 days)
GLB-3D-Coder is a code generation model with 4B parameters that can export 3D assets in GLB format by writing Three.js scene code.
GLB-3D-Coder is trained based on Qwen3.5-4B.
The assets generated by GLB-3D-Coder can be imported into game engines such as Godot, and the author demonstrated cases of assets already imported into Godot.
The author states that English prompts worked best in testing, and recommends using the system prompt in the model card together with a one-sentence natural language description.
The author recommends inference parameters of temperature 0.8, top-k 40, top-p 0.95, and min-p 0.
Most generations succeeded on the first attempt, but some still require retries or minor code modifications.
The current release is the first functional checkpoint, training is still ongoing, and the author expects training to be completed by the evening of the next day (UTC).
The author promises to release the complete Safetensors weights and the training dataset once training is complete.
A parameter count of 4B means the model can run on consumer-grade hardware, lowering the barrier to experimentation for individual developers.
For objects with relatively simple asset structures and clear semantics, the model performs more reliably; when facing scenes with complex descriptions and unique structures, the probability of rework increases.
All Facts (10)
GLB-3D-Coder is a code generation model with 4B parameters that can export 3D assets in GLB format by writing Three.js scene code.
50%UnverifiedGLB-3D-Coder is trained based on Qwen3.5-4B.
50%UnverifiedThe assets generated by GLB-3D-Coder can be imported into game engines such as Godot, and the author demonstrated cases of assets already imported into Godot.
50%UnverifiedThe author states that English prompts worked best in testing, and recommends using the system prompt in the model card together with a one-sentence natural language description.
50%UnverifiedThe author recommends inference parameters of temperature 0.8, top-k 40, top-p 0.95, and min-p 0.
50%UnverifiedMost generations succeeded on the first attempt, but some still require retries or minor code modifications.
50%UnverifiedThe author promises to release the complete Safetensors weights and the training dataset once training is complete.
50%UnverifiedA parameter count of 4B means the model can run on consumer-grade hardware, lowering the barrier to experimentation for individual developers.
50%UnverifiedThe current release is the first functional checkpoint, training is still ongoing, and the author expects training to be completed by the evening of the next day (UTC).
50%UnverifiedFor objects with relatively simple asset structures and clear semantics, the model performs more reliably; when facing scenes with complex descriptions and unique structures, the probability of rework increases.
50%