Optimized with my branch's custom auto-tensor-type, custom-made recipes for 3.77, 4.02 and 4.27 bpw
(element-gamma=0.25, tuned for MoE — recovers the bits that plain size-weighting over-spends on the
rarely-activated experts).
Since HF doesn't recognize custom bpw tags, I've tagged them with:
- IQ3_M: 3.77bpw
- Q3_K_M: 4.02bpw
- IQ4_XS: 4.27bpw
Note that the quant types are only aliases for the size and do not correspond to the actual quant types used.
Converted with --no-mtp, so the multi-token-prediction head is excluded — these are standard-inference GGUFs.