Dev fine-tunes Qwen3-1.7B into a 1.83GB GLB 3D asset coder that runs on CPU
ConfidentDinner6648 · reddit · 2026-11-11
After releasing a 4B GLB model, a developer fine-tuned Qwen3-1.7B into a much smaller 3D asset coder. The Q8 GGUF weighs just 1.83GB, targeting CPU-only or low-VRAM setups.
The workflow matches the 4B version: describe an asset, open the generated Three.js HTML, and click Export GLB to bring it into Godot or another editor — the jungle outpost demo shows the pipeline in action.
It's an early test checkpoint; some prompts still need retries, and the author plans to update weights once training finishes. Weights are available on Hugging Face at ccore/Qwen3-1.7B-GLB-3D-Coder.
More from coding & agent
- 30 Shopify engineers rebuilt a 300-screen app in 12 weeks with agents writing most of the code — schwentker · 2026-11-11
- Hugging Face adds decision pipeline: models output option probabilities without generating text — kalyan_kpl · 2026-11-11
- Anthropic says Claude ported itself to AMD GPUs in 3 days, writing ~100% of kernels unassisted — AccBalanced · 2026-11-11
- Dev tests Fable 5.2 + Unreal Engine to recreate golden-age competitive shooters — FanaHOVA · 2026-11-11
- Fable 5.2 builds a 3D game in 76 min for ~$42 in tokens, roughly matching Opus 5.5 — FinanceYF5 · 2026-11-11
- Procedural 3D dragon in threejs generated as pure code, demo online — majidmanzarpour · 2026-11-11