---

license: other

license_name: modified-mit

base_model: cerebras/MiniMax-M2-REAP-139B-A10B

base_model_relation: quantized

pipeline_tag: text-generation

library_name: gguf

tags: [gguf, quantization, moe, pruning]

---

MiniMax-M2-REAP-139B-A10B — IQ2_XXS GGUF

An IQ2_XXS quantization of cerebras/MiniMax-M2-REAP-139B-A10B,

which is itself an expert-pruned (REAP) compression of MiniMax-M2 from 230B down to 139B.

So this is a compression on top of a compression: 230B → 139B by expert pruning → ~31 GB at

about 2.06 bits per weight.

| | |

|---|---|

| file | cerebras_MiniMax-M2-REAP-139B-A10B-IQ2_XXS.gguf |

| size | 31.33 GiB |

| architecture | minimax-m2 (MoE, ~10B active) |

| parameters | 139.15 B |

| native context | 196,608 |

What is and isn't verified

Verified: loads under llama.cpp and generates — prompt processing and token generation both

exercised, exit code 0.

Not measured: no KLD or perplexity table. Two lossy steps stack here — REAP expert pruning

upstream, then a very aggressive 2-bit quantization — and this repo quantifies neither. That is a

real caveat, not boilerplate: at IQ2_XXS on an already-pruned model you should expect noticeable

degradation and should evaluate it on your own task before trusting it.

For a release where fidelity is measured against the unquantized reference, see

Qwen3.6-27B-GGUF.

Run it

llama-cli -m cerebras_MiniMax-M2-REAP-139B-A10B-IQ2_XXS.gguf -c 32768 -no-cnv -p "your prompt"

Credit & license

• Pruned model © Cerebras — MiniMax-M2-REAP-139B-A10B
• Base model © MiniMax AI — MiniMax-M2

Distributed under the modified MIT license of the upstream models: MIT terms, with the added

condition that commercial products above 100M monthly active users or $30M annual recurring revenue

must display "MiniMax M2" in their interface. This repo adds only the quantization.