AX-MiniMax-M3-MLX-AXQ-2bit
Experimental AXQ 2-bit MLX pack of
MiniMaxAI/MiniMax-M3 (BF16 source).
Not certified. Will not be certified in this revision. Layer-stack SSD
expert paging is too slow for practical serving. Hobby / curiosity only.
No AXQ 4-bit affine sibling. An MXFP4 SKU is a separate pack.
Config num_mtp_modules is not packaged MTP, so this leaf has no -MTP.
Language path only. Vision stays BF16; vision generate is not claimed. Quality vs BF16 was not measured.
| Item | Status |
|---|---|
Convert + ax_expert_stream.json |
Factory job on df-macstudio-m2 |
| Convert git SHA | 2f4e4e49b82463ac9e146090020a9565c8583253 |
| License | Upstream MiniMax LICENSE copied into the pack |
| Stream | --expert-stream required |