huggingface.co web signal

CCORE Open-Sources Qwen3.5-4B-3D-Coder for Three.js Scenes

TL;DR

  • CCORE released Qwen3.5-4B-3D-Coder, an Apache-2.0 fine-tune of Qwen/Qwen3.5-4B trained on 4,146 Three.js scene examples with full thinking traces.
  • The model ships as FP16 safetensors plus Q4_K_M and Q8_0 GGUF quantizations, with a recommended temperature of 0.6-0.8.
  • LM Studio users must disable multi-token prediction, which is on by default, for the model to work properly.

CCORE has released Qwen3.5-4B-3D-Coder, a 4B-parameter fine-tune of Qwen/Qwen3.5-4B shipped under Apache-2.0 and trained on 4,146 scene examples. The model card frames it as "an experiment exploring the feasibility of small language models for general-purpose coding and high-quality 3D scene generation," with "a particular focus on creating interactive 3D scenes with Three.js."

The weights are available as FP16 safetensors plus Q4_K_M and Q8_0 GGUF quantizations; recommended temperature is 0.6-0.8. The included system prompt pins outputs to Three.js 0.160.0 via a single ES-module CDN URL and bans addons or legacy examples paths. LM Studio users get a flat warning: "If you're running this model in LM Studio, disable multi-token prediction (MTP) for the model to work properly. LM Studio enables this option by default."

CCORE grades the work candidly. "One-shot success has improved substantially, along with overall scene quality. Some scenes can still require another attempt or a correction," the author writes, noting that most observed bugs are duplicate variable names rather than major structural problems. The drop lands among 357 open-source releases we've logged in the last 90 days.