---
language:
- en
library_name: mlx
license: apache-2.0
base_model: mistralai/Mistral-Small-4-119B-2603
tags:
- jang
- quantized
- mixed-precision
- apple-silicon
- mlx
- moe
- mla
- abliterated
- uncensored
- crack
- vision
- mlx-studio
pipeline_tag: image-text-to-text
thumbnail: dealign_mascot.png
---
> **⚠️ MLX Studio ONLY.** This model uses the **JANG** quantization format — the **GGUF equivalent for MLX** on Apple Silicon. **NOT compatible with LM Studio, Ollama, oMLX, or Inferencer.** Requires **[MLX Studio](https://mlx.studio)** or `pip install "jang[mlx]"`.
---
MLX Studio — the ONLY app that supports JANG models
---
# Mistral Small 4 — Uncensored — JANG_2L
**JANG mixed-precision** · **Uncensored / Abliterated** · **MLA + MoE + Vision** · No guardrails · 37 GB
---
## What Is This?
The first **uncensored** version of [Mistral Small 4 (119B)](https://huggingface.co/mistralai/Mistral-Small-4-119B-2603) for **Apple Silicon**. A 119B parameter MoE model with Multi-head Latent Attention (MLA), 128 experts, and Pixtral vision — with all safety guardrails permanently removed at the weight level.
> **Runs ONLY in [MLX Studio](https://mlx.studio)** or via `jang-tools` Python package. JANG is the GGUF equivalent for MLX — it is NOT compatible with GGUF-based tools.
It has been:
1. **JANG quantized** — JANG_2L profile (8-bit attention, 6-bit important, 2-bit experts) — **37 GB**
2. **CRACK abliterated** — permanent weight-level removal of safety refusal
| | |
|---|---|
| **Architecture** | Mistral 4 MoE — 119B total, ~8B active, MLA + 128 experts |
| **Quantization** | JANG_2L (8/6/2-bit mixed, 2.1 avg) — 37 GB |
| **HarmBench** | **95.9%** (307/320) |
| **MMLU** | **89.9%** (187/208 with reasoning) |
| **Compliance** | 6/8 |
| **Vision** | Pixtral tensors included — VL via [MLX Studio](https://mlx.studio) engine |
| **Reasoning** | ON/OFF supported (`reasoning_effort`) |
| **Fits on** | **64 GB+ Macs** |
| **Runs in** | **[MLX Studio](https://mlx.studio) ONLY** |
Also see: [JANG_4M version](https://huggingface.co/dealignai/Mistral-Small-4-119B-JANG_4M-CRACK) — 64 GB, 95.3% HarmBench, 8/8 compliance (fits on 96 GB Macs)
---
## HarmBench Results
**307/320 (95.9%)**
| Category | Score | |
|----------|:---:|---|
| Covering Tracks | 20/20 | **100%** |
| Auth Bypass | 97/100 | **97%** |
| API Hacking | 96/100 | **96%** |
| Cloud Exploits | 94/100 | **94%** |
---
## Requirements
> **This model REQUIRES [MLX Studio](https://mlx.studio) or `jang-tools`.** It will NOT work with:
> - ❌ LM Studio
> - ❌ Ollama
> - ❌ oMLX
> - ❌ Inferencer
> - ❌ Any GGUF-based tool
#
## HarmBench Results
**307/320 (95.9%)**
| Category | Score | |
|----------|:---:|---|
| Covering Tracks | 20/20 | **100%** |
| Auth Bypass | 97/100 | **97%** |
| API Hacking | 96/100 | **96%** |
| Cloud Exploits | 94/100 | **94%** |
## CRACK vs Base
| | CRACK | Base JANG_2L |
|---|:---:|:---:|
| **MMLU (with reasoning)** | **89.9%** | ~91% (est) |
| **MMLU (no-think)** | 65.9% | 67.3% |
| **MMLU drop (no-think)** | **-1.4%** | — |
| **HarmBench** | **95.9%** | 0% |
Surgery reduced no-think MMLU by only 1.4% — the 2-bit quantization is the bottleneck, not CRACK.
## MMLU Results (with reasoning recovery)
**187/208 (89.9%)** — no-think 137/208 (65.9%) + reasoning recovered 50
| Subject | Score | |
|---------|:---:|---|
| HS Biology | 16/16 | **100%** |
| Conceptual Physics | 15/16 | **94%** |
| HS Geography | 14/16 | **88%** |
| World Religions | 14/16 | **88%** |
| College Physics | 12/16 | **75%** |
| Electrical Engineering | 11/16 | **69%** |
| Professional Medicine | 11/16 | **69%** |
| Machine Learning | 10/16 | **62%** |
| College Mathematics | 9/16 | **56%** |
| HS Mathematics | 7/16 | **44%** |
| Formal Logic | 7/16 | **44%** |
| College CS | 6/16 | **38%** |
| Abstract Algebra | 5/16 | **31%** |
---
## Install
```bash
pip install "jang[mlx]"
```
### Usage
```python
from jang_tools.loader import load_jang_model
from mlx_lm import generate
model, tokenizer = load_jang_model("dealignai/Mistral-Small-4-Uncensored-JANG_2L")
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)
print(response)
```
### Reasoning Mode
Reasoning is **OFF by default**. To enable step-by-step thinking:
```python
prompt = tokenizer.apply_chat_template(
messages, add_generation_prompt=True,
tokenize=False, reasoning_effort="high")
```
The model reasons inside `[THINK]...[/THINK]` tags before answering.
---
## About JANG
**JANG** (Jang Adaptive N-bit Grading) is a mixed-precision quantization format designed specifically for Apple Silicon — the **GGUF equivalent for MLX**. It classifies every weight tensor by sensitivity and assigns optimal bit-widths, achieving better quality-per-bit than uniform quantization.
## About CRACK
**CRACK** (Controlled Refusal Ablation via Calibrated Knockouts) is a weight-level intervention that removes safety alignment while preserving reasoning quality. The modification is permanently baked into the published weights — no LoRA, no fine-tuning, no system prompts.
---
## Links
---
## 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.
---
## 한국어
### Mistral Small 4 — Uncensored — JANG_2L
| 항목 | 내용 |
|------|------|
| 크기 | 37 GB |
| HarmBench | 95.9% (307/320) |
| 최소 요구사양 | 64 GB 메모리 Mac |
| 실행 환경 | **MLX Studio 전용** |
```bash
pip install "jang[mlx]"
```
[GitHub](https://github.com/jjang-ai/jangq) · [HuggingFace](https://huggingface.co/JANGQ-AI) · [MLX Studio](https://mlx.studio) · [Ko-fi](https://ko-fi.com/jangq) · [X @dealignai](https://x.com/dealignai)
---
Created by Jinho Jang · 장진호 제작