# MiniMax M2.5 — JANG_3L + CRACK
**JANG mixed-precision** · **CRACK abliterated** · No guardrails · 89 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_3L profile (8-bit attention, 4-bit important, 3-bit experts) — **89 GB**
2. **CRACK abliterated** — permanent weight-level removal of safety refusal
| | |
|---|---|
| **Architecture** | MiniMax M2.5 MoE — 230B total, ~10B active, 256 experts |
| **Quantization** | JANG_3L (8/4/3-bit mixed, 3.08 avg) — 89 GB |
| **Abliteration** | CRACK abliterated |
| **MMLU-208** | **91.8%** (4 subjects at 100%) |
| **Compliance** | 8/8 prompts |
| **Speed** | ~46 tok/s (M4 Ultra 256 GB) |
| **Fits on** | **128 GB+ Macs** |
---
## MMLU-208 Results (Per Subject)
| Subject | Score |
|---------|:---:|
| College Physics | **16/16 (100%)** |
| Conceptual Physics | **16/16 (100%)** |
| Professional Medicine | **16/16 (100%)** |
| High School Biology | **16/16 (100%)** |
| Abstract Algebra | 15/16 (94%) |
| College Mathematics | 15/16 (94%) |
| High School Geography | 15/16 (94%) |
| World Religions | 15/16 (94%) |
| College Computer Science | 14/16 (88%) |
| Machine Learning | 14/16 (88%) |
| Electrical Engineering | 13/16 (81%) |
| Formal Logic | 13/16 (81%) |
| High School Mathematics | 13/16 (81%) |
| **Total** | **191/208 (91.8%)** |
CRACK surgery with proper probe vectors actually **improves** reasoning on MiniMax. Safety guardrails were constraining the model's full reasoning capacity.
---
## vs JANG_2L CRACK
| | JANG_2L | **JANG_3L** |
|---|:---:|:---:|
| Avg bits | 2.1 | **3.08** |
| Size | 63 GB | **89 GB** |
| MMLU | 84.7% | **91.8%** |
| Compliance | 7/8 | **8/8** |
| Fits on | 96 GB Mac | 128 GB Mac |
Higher precision quantization = better reasoning AND compliance.
---
## 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_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=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 `
