MiMo-V2.6-Pro-MOPD

Run with https://llama.app

llama serve -hf pcuenq/MiMo-V2.6-Pro-MOPD-GGUF

Source models

• https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-MOPD

Quants

• MXFP4
• Q3_K
• Q2_K

Notes

• The MXFP4 quant keeps the routed experts at their native MXFP4 precision.
• The Q3_K quant keeps the expert down projections at MXFP4, and quantizes the gate/up projections to Q3_K.
• The Q2_K quant keeps the expert down projections at MXFP4, and quantizes the gate/up projections to Q2_K.
• Includes MTP sidecars (Q4_0 and Q8_0) for speculative decoding (--mtp).
• Includes a DFlash drafter sidecar (BF16 and Q8_0) for speculative decoding, converted from the dflash/ subdirectory of the source repo.
• Includes a mmproj adapter (BF16 and Q8_0) for the vision and audio encoders.
• The Q3_K and Q2_K expert gate/up tensors are calibrated with the imatrix from https://huggingface.co/AesSedai/MiMo-V2.6-Pro-MOPD-GGUF

This model was manually converted by pcuenq following the approach used in ggml-org/MiMo-V2.6-Flash-MOPD-GGUF.