Users are questioning ISTA's GSQ-RCO 2-bit quantized GGUF versions of DeepSeek-V4.1-Flash, noting that reported benchmarks (90%+ saturated, AIME25 at 100%) may not reveal quality degradation. The post seeks community validation on whether IQ2_XXS models match base model quality and requests measurements of Kullback-Leibler divergence (KLD) or perplexity (PPL) to better assess performance at approximately 2 bits per weight.