license: other
license_name: other
license_link: https://github.com/MiniMax-AI/MiniMax-M2.7/blob/main/LICENSE
base_model:
- llmfan46/MiniMax-M2.7-BF16-ultra-uncensored-heretic
library_name: mlx
pipeline_tag: text-generation
tags: - mlx
- omlx
- oq
- oq4
- quantized
- minimax_m2
- heretic
- uncensored
- decensored
- abliterated
- ara
MiniMax-M2.7-ultra-uncensored-heretic-oQ4-MLX
This repository contains an oMLX oQ4 mixed-precision MLX quantization ofllmfan46/MiniMax-M2.7-BF16-ultra-uncensored-heretic.
This build follows the artifact lineage fromllmfan46/MiniMax-M2.7-ultra-uncensored-heretic-GGUF.
oMLX oQ quantization operates on MLX/safetensors checkpoints rather than GGUF
files, so this build uses the corresponding BF16 safetensors checkpoint and
excludes the existing GGUF quantizations.
Quantization
| Field | Value |
|---|---|
| Method | oMLX oQ mixed-precision MLX |
| Quantization | oQ4 |
| Model type | minimax_m2 |
| Group size | 64 |
| Quantization mode | affine |
| Effective plan | 4.57 bpw |
| Layer policy entries | 250 |
| Output shards | 24 safetensors |
| Output size | 121.7 GiB |
Usage
huggingface-cli download dawncr0w/MiniMax-M2.7-ultra-uncensored-heretic-oQ4-MLX \
--local-dir MiniMax-M2.7-ultra-uncensored-heretic-oQ4-MLX
Then load it with an MLX-LM/oMLX runtime that supports minimax_m2:
python -m mlx_lm.generate \
--model MiniMax-M2.7-ultra-uncensored-heretic-oQ4-MLX \
--prompt "Write a short greeting." \
--max-tokens 64
Validation
Local validation completed with the bundled oMLX runtime on macOS:
model discovery: passed
model type: minimax_m2
quantization: bits=4, group_size=64, mode=affine
shards: 24
Source
- BF16 source checkpoint:
llmfan46/MiniMax-M2.7-BF16-ultra-uncensored-heretic - GGUF reference requested for this build:
llmfan46/MiniMax-M2.7-ultra-uncensored-heretic-GGUF - Sensitivity proxy used during quantization:
cookietimeh/MiniMax-M2.7-BF16-ultra-uncensored-heretic-mlx-4Bit - Quantization tool: oMLX oQ
Upstream model card license tag: other.