GSQ + RCO
by IST Austria DAS Lab · Technique · Text & code
ISTA's DASLab (GPTQ authors) released GSQ and RCO quantization methods, shrinking Qwen3.8-27B to 8-12 GB
Evidence
- Sep 7GSQ picks per-weight bit depth (2-4 bit); RCO spreads a fixed size budget across tensors, giving more bits to sensitive ones
Alternatives
Made for GSQ + RCO