Too good to be true? ISTA's GSQ-RCO 2-bit DeepSeek-V4.1-Flash scoring the same as the base model
retained text
Has anyone tried ISTA's GSQ-RCO GGUFs for DeepSeek-V4.1-Flash? Does the quality feel as good as their benchmarks show (IQ2\_XXS matching the base)? Their benchmarks are saturated (90%+, and AIME25 is 100 for every version), so they probably can't show the degradation, and I couldn't find any KLD for GSQ-RCO models. Has anyone measured KLD or PPL? What KLD would you expect at \~2 bpw? [https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF](https://huggingface.co/ISTA-DASLab/DeepSeek-V4.1-Flash-1017-GSQ-RCO-GGUF) (English isn't my first language, I used an LLM to help translate this post.)
document fetched 2026-10-26 00:38 · canonical https://reddit.com/r/LocalLLaMA/comments/1xes5x0/too_good_to_be_true_istas_gsqrco_2bit