This repo contains specialized MoE-quants for MiniMax-M2.5. The idea being that given the huge size of the FFN tensors compared to the rest of the tensors in the model, it should be possible to achieve a better quality while keeping the overall size of the entire model smaller compared to a similar naive quantization. To that end, the quantization type default is kept in high quality and the FFN UP + FFN GATE tensors are quanted down along with the FFN DOWN tensors.
Quant
Size
Mixture
PPL
1-(Mean PPL(Q)/PPL(base))
KLD
Q5_K_M
157.23 GiB (5.91 BPW)
Q8_0 / Q5_K / Q5_K / Q6_K
7.126261 ± 0.115850
+0.5877%
0.023465 ± 0.001079
Q4_K_M
130.52 GiB (4.90 BPW)
Q8_0 / Q4_K / Q4_K / Q5_K
7.173459 ± 0.116673
+1.2462%
0.041269 ± 0.001426
IQ4_XS
101.10 GiB (3.80 BPW)
Q8_0 / IQ3_S / IQ3_S / IQ4_XS
7.513587 ± 0.122746
+6.0549%
0.095077 ± 0.002168
IQ3_S
78.76 GiB (2.96 BPW)
Q8_0 / IQ2_S / IQ2_S / IQ3_S
8.284882 ± 0.135705
+16.9418%
0.244096 ± 0.004148
Provided here as well as a couple of graphs showing the Pareto frontier for KLD and PPL for my quants vs Unsloth.
Full graphs of all of the quants are available in the kld_data directory, as well as the raw data broken down per quant as well as a CSV with the collated data.
While the PPL between the quant methods is similar, I feel like the KLD of the quants provided here are slightly better and that these quants will offer better long context performance due to keeping the default type as Q8_0. This comes with a slight performance penalty in PP / TG due to the higher quality quantization but I think the tradeoff is worthwhile.
