AesSedai/MiMo-V2.5-GGUF

Notes

• 04/28/26: While this model should run on the llama.cpp master branch, there was a small change to the inference code to support the attention_value_scale parameter. For the best accuacy/performance, I recommend pulling and compiling from this PR branch: https://github.com/ggml-org/llama.cpp/pull/22493.

Model

This is a text-only GGUF quantization of XiaomiMiMo/MiMo-V2.5. This means that image and audio input is not present in this GGUF, and will not be available until support is added upstream in llama.cpp.

This repo contains specialized MoE-quants for MiMo-V2.5. 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
Q8_0	305.68 GiB (8.50 BPW)	Q8_0	5.135221 ± 0.030263	+0.1229%	0.012455 ± 0.000173
Q5_K_M	212.42 GiB (5.91 BPW)	Q8_0 / Q5_K / Q5_K / Q6_K	5.142293 ± 0.030329	+0.2608%	0.014906 ± 0.000230
Q4_K_M	176.70 GiB (4.92 BPW)	Q8_0 / Q4_K / Q4_K / Q5_K	5.204791 ± 0.030827	+1.4793%	0.020743 ± 0.000168
IQ4_XS	136.78 GiB (3.80 BPW)	Q8_0 / IQ3_S / IQ3_S / IQ4_XS	5.270911 ± 0.031163	+2.7685%	0.041606 ± 0.000299
IQ3_S	105.33 GiB (2.93 BPW)	Q6_K / IQ2_S / IQ2_S / IQ3_S	5.547074 ± 0.033209	+8.1529%	0.092593 ± 0.000568