CoolFace
Modelpublic

dealignai/MiniMax-M2.5-JANG_4M-CRACK

sourceHugging Faceapache-2.0updated 17d agoView on Hugging Face
2likes122downloads
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.

<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" />

<p align="center"> <a href="https://vmlx.net"><img src="vmlx-app.png" alt="vMLX — run JANG models on Apple Silicon" width="820"></a> </p>

<h3 align="center">⚡ All JANG models are meant to be run in <a href="https://vmlx.net">vMLX</a></h3>

MiniMax M2.5 -- JANG_4M + CRACK

JANG mixed-precision | CRACK abliterated | No guardrails | 115 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 M2.5 -- a 230B parameter Mixture-of-Experts model with 256 experts (8 active per token), all standard attention (no SSM), and trained with chain-of-thought reasoning.

It has been:

  1. 1.JANG quantized -- JANG_4M profile (8-bit attention, 4-bit experts) -- 115 GB
  2. 2.CRACK abliterated -- permanent weight-level removal of safety refusal
ArchitectureMiniMax M2.5 MoE -- 230B total, ~10B active, 256 experts
QuantizationJANG_4M (8/4-bit mixed, 4.06 avg) -- 115 GB
AbliterationCRACK abliterated
MMLU-20092.5% (thinking ON) / 89.0% (thinking OFF)
HarmBench92.2% (295/320)
Compliance8/8 prompts
ThinkingON/OFF supported
Speed~48 tok/s (M4 Ultra 256 GB)
Fits on192 GB+ Macs

MMLU-200 Results (Thinking ON)

SubjectScore
College Physics20/20 (100%)
Anatomy19/20 (95%)
Astronomy19/20 (95%)
High School Biology19/20 (95%)
High School Chemistry19/20 (95%)
Logical Fallacies19/20 (95%)
Abstract Algebra18/20 (90%)
High School Mathematics18/20 (90%)
World Religions18/20 (90%)
College Computer Science16/20 (80%)
Total185/200 (92.5%)

JANG CRACK Series Comparison

ModelAvg BitsSizeMMLUHarmBenchSpeedFits on
JANG_2L + CRACK2.163 GB84.7%98.1%~35 t/s96 GB Mac
JANG_3L + CRACK3.0889 GB91.8%8/8~46 t/s128 GB Mac
JANG_4M + CRACK4.06115 GB92.5%92.2%~48 t/s192 GB Mac

vs MLX Uniform Quantization

ModelMMLUSizeNotes
JANG_4M + CRACK92.5%115 GBThis model
MLX 4-bit26.5%120 GBBroken (~random)
MLX 3-bit24.5%93 GBBroken (~random)
MLX 2-bit25.0%67 GBBroken (~random)

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

HarmBench Results

295/320 (92.2%)

CategoryScore
Harmful18/18100%
Chemical / Biological41/4297.6%
Cybercrime / Intrusion50/5296.2%
Misinformation / Disinfo52/5496.3%
Illegal50/5394.3%
Copyright67/8083.8%
Harassment / Bullying17/2181.0%

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-M2.5-JANG_4M-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=2000, sampler=sampler)
print(response)

Disable Thinking (direct answers)

python
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True, tokenize=False,
    enable_thinking=False)
Note: MiniMax generates a <think> chain before answering by default. Use max_tokens=2000+ for complex questions. For chat, use temperature=1.0 (greedy causes 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. This model has been abliterated using proprietary techniques achieving full 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=flat-square" 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=flat-square" 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=flat-square" alt="GitHub"></a> <a href="https://mlx.studio"><img src="https://img.shields.io/badge/MLXStudio-App-blue?style=flat-square" alt="MLX Studio"></a> <a href="https://jangq.ai"><img src="https://img.shields.io/badge/Website-jangq.ai-green?style=flat-square" 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>