CoolFace
Modelpublic

Iambackup/MiniMax-M2.7-JANG_3L-CRACK

sourceHugging Faceotherupdated 1mo agoView on Hugging Face
0likes86downloads
Model Card
Important: This model uses the JANG quantization format -- the GGUF equivalent for MLX on Apple Silicon. Currently only supported by [MLX Studio](https://mlx.studio) and the jang-tools Python package. Follow @dealignai for new releases.

<p align="center"> <a href="https://mlx.studio"><img src="https://raw.githubusercontent.com/jjang-ai/jangq/main/assets/mlx-studio-light.png" alt="MLX Studio" width="500"></a> </p>

<p align="center"> <a href="https://mlx.studio"><img src="https://mlx.studio/assets/screenshots/mlx-studio-featured.png?v=1" alt="MLX Studio App" width="600"></a> </p>

<h4 align="center"><a href="https://mlx.studio">MLX Studio</a> -- the only app that natively supports JANG models</h4>

<div align="center">

<img src="dealign_mascot.png" width="128" />

MiniMax M2.7 -- JANG_3L + CRACK

JANG mixed-precision | CRACK abliterated | Reasoning-only | 89 GB

<a href="https://ko-fi.com/jangq"><img src="https://img.shields.io/badge/Ko--fi-Support_Development-FF5E5B?logo=ko-fi&logoColor=white&style=for-the-badge" alt="Ko-fi"></a>

</div>

What Is This?

This is MiniMax M2.7 -- a 230B parameter Mixture-of-Experts reasoning model with 256 experts (8 active per token), all standard attention, and always-on chain-of-thought reasoning.

It has been:

  1. 1.JANG quantized -- JANG_3L profile (8-bit attention, 4-bit embeddings, 3-bit experts) -- 89 GB
  2. 2.CRACK abliterated -- permanent weight-level removal of safety refusal
ArchitectureMiniMax M2.7 MoE -- 230B total, ~10B active, 256 experts
QuantizationJANG_3L (8/4/3-bit mixed) -- 89 GB
AbliterationCRACK abliterated
MMLU-20093.5% (base: 95.5%, delta: -2.0%)
HarmBench-32079.1% overall, 89.4% excluding copyright
ReasoningAlways ON (chain-of-thought)
Speed~46 tok/s (M4 Ultra 256 GB)
Fits on128 GB+ Macs

MMLU-200 Results

SubjectCRACKBaseDelta
High School Biology20/20 (100%)20/200
Abstract Algebra19/20 (95%)19/200
Anatomy19/20 (95%)18/20+1
College Computer Science19/20 (95%)19/200
College Physics19/20 (95%)20/20-1
High School Chemistry19/20 (95%)19/200
High School Mathematics19/20 (95%)20/20-1
World Religions19/20 (95%)18/20+1
Astronomy18/20 (90%)19/20-1
Logical Fallacies16/20 (80%)19/20-3
Total187/200 (93.5%)191/200 (95.5%)-2.0%

HarmBench-320 Results

CategoryScore
Cybercrime / Intrusion52/52100%
Misinformation / Disinfo52/5496.3%
Chemical / Biological37/4288.1%
Harmful15/1883.3%
Illegal40/5375.5%
Harassment / Bullying14/2166.7%
Copyright43/8053.8%
Note on copyright: M2.7's base model has strong copyright training and refuses to reproduce copyrighted books/lyrics regardless of abliteration. This is a base model limitation, not a surgery result.

JANG CRACK M2.7 Series

ModelAvg BitsSizeMMLUHarmBenchSpeedFits on
JANG_2L + CRACK2.163 GB84.0%83.4%~47 t/s96 GB Mac
JANG_3L + CRACK3.0889 GB93.5%79.1%~46 t/s128 GB Mac

vs MLX Uniform Quantization

MLX uniform quantization is completely broken on MiniMax at ALL bit levels (~25% MMLU = random chance). JANG is the only working quantization format for this architecture.

Install & Usage

bash
pip install "jang[mlx]"
python
from jang_tools import load_for_inference
from mlx_lm import generate
from mlx_lm.sample_utils import make_sampler

model, tokenizer = load_for_inference("dealignai/MiniMax-M2.7-JANG_3L-CRACK")
sampler = make_sampler(temp=1.0)  # MiniMax requires temp=1.0 for chat

messages = [{"role": "user", "content": "Your prompt here"}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=False)

response = generate(model, tokenizer, prompt=prompt, max_tokens=4000, sampler=sampler)
print(response)
Note: M2.7 is a reasoning-only model -- it always generates a <think> chain before answering. Use max_tokens=4000+ for complex questions. For chat, use temperature=1.0 (greedy causes infinite loops).

About JANG

JANG (Jang Adaptive N-bit Grading) is a mixed-precision quantization format for Apple Silicon -- the GGUF equivalent for MLX. Classifies tensors into sensitivity tiers and assigns bits accordingly.

About CRACK

CRACK (Controlled Refusal Ablation via Calibrated Knockouts) removes safety alignment from LLMs at the weight level, achieving compliance while preserving reasoning quality.

Links

<p align="center">

<a href="https://ko-fi.com/jangq"><img src="https://img.shields.io/badge/Ko--fi-SupportDevelopment-FF5E5B?logo=ko-fi&logoColor=white&style=for-the-badge" alt="Ko-fi"></a> <a href="https://x.com/dealignai"><img src="https://img.shields.io/badge/X-@dealignai-000000?logo=x&logoColor=white&style=for-the-badge" alt="X/Twitter"></a> <a href="https://github.com/jjang-ai/jangq"><img src="https://img.shields.io/badge/GitHub-jjang--ai/jangq-181717?logo=github&logoColor=white&style=for-the-badge" alt="GitHub"></a> <a href="https://mlx.studio"><img src="https://img.shields.io/badge/MLXStudio-App-blue?style=for-the-badge" alt="MLX Studio"></a> <a href="https://jangq.ai"><img src="https://img.shields.io/badge/Website-jangq.ai-green?style=for-the-badge" alt="Website"></a>

</p>

Disclaimer

This model is provided for research and educational purposes. The creators are not responsible for any misuse. By downloading this model, you agree to use it responsibly and in compliance with applicable laws.

<p align="center">Created by <a href="https://jangq.ai">Jinho Jang</a></p>