4B 3D coding model update: much better one-shot results and visual quality
retained text
Quick update: I've released the latest version of my **4B coding model**, including the **full FP16 weights**, and the **training dataset is now public!** One-shot results and visual quality are much better in my latest tests, and training is nearly finished. Some scenes still need another attempt or a correction, but the progress has been encouraging. **I recommend starting with Q8**, which is the version I’ve tested. Q4 is also available, but I haven’t tested this release yet. **Q8 settings:** temperature **0.6**, top-k **20**, top-p **0.95**. I’ve also had good results generating **3D scenes from image references**, even though the vision encoder wasn’t included in this training run. [**Updated Q4 and Q8 on Hugging Face**](https://huggingface.co/ccore/Qwen3.5-4B-3D-Coder) **If you downloaded an earlier version, delete it and download it again in LM Studio or whichever app you use. The filenames haven’t changed, so make sure you’re loading the updated files.** In LM Studio, also **disable MTP**. **Dataset released:** [**CCORE Three.js 3D Scenes**](https://huggingface.co/datasets/ccore/ThreeJS-3D-Scenes) It contains **4,146 scene examples** — 3,933 training and 213 validation — with prompts, full thinking traces, complete HTML/JavaScript responses, complexity levels, and optional rendered reference images. It’s available under the **MIT License** for anyone who wants to train other models. As promised, the **full FP16 weights have now been released!** Thanks everyone for the support, feedback, and testing! It’s been great seeing the community try it out, and I’d love to see what you build with the dataset.
document fetched 2026-10-29 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xhfwui/4b_3d_coding_model_update_much_better_oneshot