Deceptive model quantization from AtomicChat?
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
I kept seeing guys in this sub saying how AtomicChat's Qwen3.8-Flash-Next quant is so good, fits in their machine when unsloth's can't, runs faster than other quants etc, so I went check out what's happening there.
First thing I noticed was that AtomicChat's Q4_K_M quant is suspiciously small when the ngram table is removed (only ~56GB), it seems like most of the tensors in this quant are IQ2_S instead of the usual Q4_K, Q5_K and Q6_K that you usually find in Q4_K_M quants, the GGUF filetype metadata also says IQ2_S instead of Q4_K_M. In their model card, their Q4_K_M also has suspiciously high KLD (0.084).
It seems pretty obvious to me that they're pretending a IQ2_S quant as a Q4_K_M, but at the same time I'm genuinely not sure because it can't be only me who found this right? How can nobody be pointing this out? Am I missing something or what may they be doing?
Their HF repo ID: AtomicChat/Qwen3.8-Flash-Next-GGUF