MiniMax-M3.1-Flash abliterated Q3-4

Public copy of a mixed GGUF made so the file can be pulled onto another machine. Q3-4 names the two big expert precisions: routed gate and up are Q3_K, and the down-experts are 4-bit NVFP4 except the abliterated ones, which stay Q6_K. The whole file is 4.31 bits per weight. The MiniMax Community License applies, including its non-commercial terms and prohibited-uses appendix.

Built with MiniMax M3.1-Flash. This is a requantization of operationrange/MiniMax-M3.1-Flash-abliterated-GGUF revision ef198a171043c687a768a4179917ab6547672583. No new abliteration was run. The edited tensors are the ones that repo already stored as Q8_0.

File Size
MiniMax-M3.1-Flash-abliterated-fit256.gguf 214.03 GiB

Tensor mix, from that abliterated Q8 file:

  • blk.16–blk.42 ffn_down_exps (the abliterated experts): Q6_K
  • other routed down-experts: original NVFP4, copied
  • routed gate and up experts: Q3_K
  • attention Q/K/output, shared expert, dense lead FFN: Q8_0
  • attention V, embeddings, output head: BF16
  • norms, router, indexer, and per-expert .scale tensors: copied

The per-expert scale tensors are required. DwarfStar multiplies gate and up by that scale before SwiGLU, and down by that scale before the routed reduction. A runtime that skips the scales prints token soup. Homebrew llama.cpp 0.7.0 does that on this file.

hf download choppedgarlic/MiniMax-M3.1-Flash-abliterated-Q3-4 --local-dir ./minimax-q3-4

The model will comply with requests the original refuses.

Measured on a 256 GB M3 Ultra

Resident in DwarfStar, context 65536. The weights file in this repo is the one that was measured. It was not rebuilt for these numbers.

tok/s
Decode, no drafter 18.95
Decode, DFlash2, predictable integers, temperature 0 31
Prefill, 7472 tokens 74
Thinking on, sampled about 20

The 31 tok/s figure is the greedy integer bench, at 94.9% DFlash2 acceptance. A normal thinking reply stays near 20 tok/s, because sampled turns commit one token at a time. Keeping argmax drafts during temperature-1 thinking locked repetitions.

Write-up, including the quant recipe and the gibberish notes: jjakemaness/dwarfstar-minimax-m31. Engine branch: minimax-m31-m3ultra. Pull request: antirez/ds4#1320, stacked on #1287.

Downloads last month
11,547
GGUF
Model size
426B params
Architecture
minimax-m3
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for choppedgarlic/MiniMax-M3.1-Flash-abliterated-Q3-4

Quantized
(42)
this model