MXFP4_MOE quantization of Vortex5/Phoenix-X-26B-A4B with more aggressive compression than standard mxfp4_moe. Local attention at IQ4_XS, global attention at Q5_K, attn_output in global attention layers in Q6_K.

imatrix generated by alexokita

Disclosure

My only contribution is compute. This is not my merge. Have fun.

Koshkasa/Vortex5_Phoenix-X-26B-A4B-MXFP4_MOE-GGUF

What's that?

Disclosure

Model card incomplete. Tests and comparisons may be uploaded at a later date.

Koshkasa/
Vortex5_Phoenix-X-26B-A4B-MXFP4_MOE-GGUF