Plan for roughly 9 GB of VRAM on the 4B model at standard precision — that is the figure circulating in community setup write-ups since launch. Quantized (compressed) weights trade some output quality for a smaller footprint, and the 1.7B GLB variant exists precisely for machines that can't feed the 4B model.
| Variant | Community-reported ballpark | Note |
|---|---|---|
| Qwen 3.5 4B 3D Coder | ~9 GB | Standard precision; comfortable on 12 GB+ cards |
| Qwen 3.5 4B GLB 3D Coder | ~9 GB class | Same family, different output format; verify per repo |
| Qwen 3 1.7B GLB Coder | meaningfully lower | Built for smaller GPUs; quality trade-off |
| Quantized community builds | varies | Smaller footprint, some quality cost; follow the quantizer's notes |
CUDA out of memory, see troubleshooting for the usual fixes.On NVIDIA: run nvidia-smi -l 1 in a terminal while generating and watch the memory column peak. On Windows, Task Manager → Performance → GPU shows shared vs dedicated memory. Record your peak, add headroom, and you have your real number — far more useful than any guide's estimate, including this one.
Community write-ups report roughly 9 GB of VRAM for the 4B variant in typical use, with the 1.7B GLB variant targeting smaller cards. Exact usage depends on backend, precision and quantization — verify against the official repo.
Users report workable results with quantized weights or the 1.7B GLB variant on 8 GB cards, with longer load times and some quality trade-off. Close other GPU apps first.