A 4B open-weight coding model (Qwen3.5-4B-GLB-3D-Coder) that generates Three.js scenes with GLB export for creating 3D game assets usable in engines like Godot is being released as an initial checkpoint with full weights and training data to follow.
Hey everyone! After sharing my previous 3D coding model, a lot of you asked about using it to generate assets for your games. That sounded like a fun direction to explore, so I started training a version focused on exactly that. The model writes Three.js scenes with GLB export, so you can take the generated assets into editors like Godot. The video shows some examples I've been testing, including assets imported into Godot. You don't need complicated prompts. Use the system prompt in the model card, then try a simple request like “Create a mana potion” or “Build a weathered frontier medic station.” English prompts have worked best in my tests. This is the first functional checkpoint, and training is still running. I'm sharing it now so people can start experimenting. Most generations have worked on the first attempt in my tests, but some still need a retry or a small code fix. I'm hoping later checkpoints will improve on that. My current settings: temperature 0.8, top-k 40, top-p 0.95, min-p 0. Feel free to experiment with other combinations! Training is expected to finish tomorrow evening (UTC), depending on progress. Just like with the previous model, I'll release the full Safetensors weights and the training dataset once it finishes, so others can build on this and experiment with training their own models. Model: Qwen3.5-4B-GLB-3D-Coder Thanks for all the feedback on the previous release ... it directly shaped this one. I'd love to see what you make with it and hear what works and what breaks!
The Qwen3.8 27b AI model is highlighted for its ability to handle 3D models and GLB files in game development repositories, though it cannot generate them directly.
Bad Theory Labs released BTL-3, a 27B open-weight agentic coding and tool-use model that fits in 8.39GB via custom quantization, achieving 92.2% performance retention and strong benchmarks including HumanEval 95.12% pass@1.
3DCodeVerse introduces a research ecosystem with a 2M-scale dataset of executable 3D code, a multi-agent DAG harness for automated generation and verification of 3D content, and empirical guidance showing fine-tuned open models achieve a 90% win rate against Claude Design (3D Object).