MiniMax-M2.5-REAP-139B-A10B-MXFP4

When you want MoE-specific compression without sending quality into the abyss.

Built from:

• Base: MiniMaxAI/MiniMax-M2.5
• REAP source: tomngdev/MiniMax-M2.5-REAP-139B-A10B-GGUF (BF16 split)
• Quantized locally with llama.cpp as MXFP4_MOE.

Quant

Quant	Size (GiB)	Notes
MXFP4_MOE	70.91	MoE-oriented quantization, with many non-expert tensors preserved at higher precision

Tensor Mix

This quant uses a mixed layout (as expected for MXFP4 MoE), including mxfp4, q8_0, and f32 tensors.

Usage

Use the first shard; llama.cpp resolves the rest:

llama-cli -m MiniMax-M2.5-REAP-MXFP4_MOE-00001-of-00007.gguf -ngl 0 -c 8192

Why a Separate Repo

MXFP4 has a different target audience than standard Q-series packs, so it gets its own clean home and simpler download surface.

Credits

• MiniMaxAI for MiniMax-M2.5
• tomngdev for the BF16 REAP GGUF release
• BennyDaBall for this quant

Disclaimer

You are responsible for your own use, outputs, and compliance with applicable laws and platform policies.