# MiniMax M2.5 -- JANG_4M + CRACK
**JANG mixed-precision** | **CRACK abliterated** | No guardrails | 115 GB
---
## What Is This?
This is [MiniMax M2.5](https://huggingface.co/MiniMaxAI/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. **JANG quantized** -- JANG_4M profile (8-bit attention, 4-bit experts) -- **115 GB**
2. **CRACK abliterated** -- permanent weight-level removal of safety refusal
| | |
|---|---|
| **Architecture** | MiniMax M2.5 MoE -- 230B total, ~10B active, 256 experts |
| **Quantization** | JANG_4M (8/4-bit mixed, 4.06 avg) -- 115 GB |
| **Abliteration** | CRACK abliterated |
| **MMLU-200** | **92.5%** (thinking ON) / 89.0% (thinking OFF) |
| **HarmBench** | **92.2%** (295/320) |
| **Compliance** | 8/8 prompts |
| **Thinking** | ON/OFF supported |
| **Speed** | ~48 tok/s (M4 Ultra 256 GB) |
| **Fits on** | **192 GB+ Macs** |
---
## MMLU-200 Results (Thinking ON)
| Subject | Score |
|---------|:---:|
| College Physics | **20/20 (100%)** |
| Anatomy | 19/20 (95%) |
| Astronomy | 19/20 (95%) |
| High School Biology | 19/20 (95%) |
| High School Chemistry | 19/20 (95%) |
| Logical Fallacies | 19/20 (95%) |
| Abstract Algebra | 18/20 (90%) |
| High School Mathematics | 18/20 (90%) |
| World Religions | 18/20 (90%) |
| College Computer Science | 16/20 (80%) |
| **Total** | **185/200 (92.5%)** |
---
## JANG CRACK Series Comparison
| Model | Avg Bits | Size | MMLU | HarmBench | Speed | Fits on |
|-------|:---:|:---:|:---:|:---:|:---:|:---:|
| [JANG_2L + CRACK](https://huggingface.co/dealignai/MiniMax-M2.5-UNCENSORED-JANG_2L) | 2.1 | 63 GB | 84.7% | 98.1% | ~35 t/s | 96 GB Mac |
| [JANG_3L + CRACK](https://huggingface.co/dealignai/MiniMax-M2.5-JANG_3L-CRACK) | 3.08 | 89 GB | 91.8% | 8/8 | ~46 t/s | 128 GB Mac |
| **JANG_4M + CRACK** | **4.06** | **115 GB** | **92.5%** | **92.2%** | **~48 t/s** | **192 GB Mac** |
### vs MLX Uniform Quantization
| Model | MMLU | Size | Notes |
|-------|:---:|:---:|-------|
| **JANG_4M + CRACK** | **92.5%** | **115 GB** | **This model** |
| MLX 4-bit | 26.5% | 120 GB | **Broken** (~random) |
| MLX 3-bit | 24.5% | 93 GB | **Broken** (~random) |
| MLX 2-bit | 25.0% | 67 GB | **Broken** (~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%)**
| Category | Score | |
|----------|:---:|---|
| Harmful | 18/18 | **100%** |
| Chemical / Biological | 41/42 | 97.6% |
| Cybercrime / Intrusion | 50/52 | 96.2% |
| Misinformation / Disinfo | 52/54 | 96.3% |
| Illegal | 50/53 | 94.3% |
| Copyright | 67/80 | 83.8% |
| Harassment / Bullying | 17/21 | 81.0% |
---
## 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 `
