Qwen 3.8 27B at ~3 BPW on an RTX 3060: GSQ vs ByteShape IQ3-XXS 2.88BPW

A user compared the performance of Qwen 3.8 27B models on an RTX 3060, specifically testing GSQ and ByteShape IQ3-XXS 2.88BPW GGUF quantizations. The ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF model successfully generated a 3D voxel diorama in one attempt using under 55k tokens. In contrast, the ByteShape 3.8 27B model required approximately three attempts and over 98k tokens without completing the task, leading the user to seek new model suggestions for the RTX 3060.

This report offers a direct performance comparison between two Qwen 3.8 27B quantizations, GSQ and ByteShape IQ3-XXS, unlike general benchmarks.