CCORE previews Qwen3.5-4B-3D-Coder, a 4B Three.js coding fine-tune
Original title: ccore/Qwen3.5-4B-3D-Coder — new model trending #30 on Hugging Face
The independent early preview, trending #30, refreshed Q4_K_M and Q8_0; full weights and the dataset follow after October 28.
CCORE published an early preview of Qwen3.5-4B-3D-Coder, an independent fine-tune of Qwen/Qwen3.5-4B aimed at general coding and interactive Three.js 3D scenes. It is not an official Qwen release. On October 26, 2026, the author updated the Q4_K_M and Q8_0 GGUF files, citing a layer-related fix and better one-shot scene results in the author's tests. Training is expected to finish on October 28, 2026, at 07:30 UTC, after which full FP16 weights, quantizations, the dataset, and an evaluation report are planned.
Unconfirmed: Claims of more successful one-shot scene generation come only from the author's latest tests.
- The model card says it is trending #30 on Hugging Face.
- Q8_0 GGUF updates are planned for October 26 and 27 from the latest checkpoint.
- Recommended sampling is temperature 0.6–0.8, top-p 0.90–0.95, top-k 20–40, and repeat penalty 1.0.
- LM Studio users are told to disable multi-token prediction, which is on by default.
- Scenes are meant to use Three.js 0.160.0 from cdnjs without addons.
- Organizations
- CCORE
- AI models
- Qwen3.5-4B-3D-CoderQwen3.5-4B
From the original article
4 min read at huggingface.co**An experiment exploring the feasibility of small language models for general-purpose coding and high-quality 3D scene generation. ** This is an early preview of a 4B model I am still training, with a particular focus on creating interactive 3D scenes with Three. js.
By ccore · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.