Edited - After digging into this some more and reviewing unsloth post for better understanding, the divergence APPEARS to stem from I did not use the BF16 QAT model as the "reference" model.... The QAT vs standard Q4 comparison in our benchmark is not apples-to-apples . The QAT models were evaluated against a reference they were never optimised toward. The standard Q4_0 and Q4_K_M comparison is valid. The "QAT is worse" conclusion needs a big asterisk: we can't actually ...