create a bakery display counter with a glass case, loaves of bread, and wicker baskets.

Qwen3-1.7B-GLB-3D-Coder-GGUF

CCORE GLB-3D-Coder 1.7B is an independent fine-tune of Qwen3-1.7B by CCORE (not an official Qwen release). It was trained on the CCORE ThreeJS-3D-GLB-Assets dataset to write complete standalone HTML pages that render a low-poly Three.js asset with an interactive 3D preview and an Export GLB button. The target is game assets for the Godot editor, with animation included in the exported GLB when the asset needs it. The card says the aim was a short iteration loop on modest hardware, and that Qwen3-1.7B keeps grouped-query attention across all 28 layers. The dataset covers 40 domains, from characters, creatures and bosses to props, vehicles, architecture and interactive objects. It is an early-stage release, and the card gives no benchmark numbers. The repository has GGUF builds in Q4_K_M (1.11 GB), Q8_0 (1.83 GB, the suggested starting point for CPU or low-VRAM use) and F16 (3.45 GB), all from one checkpoint last updated on November 13, 2026, plus FP16 Safetensors for Transformers. The main tests used F16 and Q8_0, and the author warns that Q4_K_M gave substantially lower-quality output. The card recommends temperature 0.8, top-k 40, top-p 0.95 and a repeat penalty of 1.0, along with a fixed system prompt. Some attempts fail to render or export and need a retry. It is released under Apache 2.0.

Model Files

llama.cpp

LLM inference in C/C++ — https://github.com/ggml-org/llama.cpp

Quantized using llama.cpp by prithivMLmods.