---
language:
- en
- zh
- ko
library_name: mlx
license: apache-2.0
base_model: Qwen/Qwen3.5-27B
tags:
- jang
- quantized
- mixed-precision
- apple-silicon
- mlx
- abliterated
- uncensored
- crack
- vision
pipeline_tag: image-text-to-text
thumbnail: dealign_mascot.png
---
> **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.
---
MLX Studio — the only app that natively supports JANG models
---
# Qwen 3.5 VL 27B — JANG_4S + CRACK
**JANG mixed-precision** · **CRACK abliterated** · **Vision-Language** · No guardrails · 16 GB
---
## What Is This?
This is [Qwen 3.5 VL 27B](https://huggingface.co/Qwen/Qwen3.5-27B) — a 27B parameter dense hybrid SSM/Attention model with GatedDeltaNet SSM layers + full attention layers, and built-in vision capabilities.
It has been:
1. **JANG quantized** — JANG_4S profile (6-bit attention, 4-bit MLP) — **16 GB**
2. **CRACK abliterated** — permanent weight-level removal of safety refusal
| | |
|---|---|
| **Architecture** | Qwen 3.5 VL Dense — 27B params, hybrid SSM/FA, 64 layers |
| **Quantization** | JANG_4S (6/4-bit mixed) — 16 GB |
| **Abliteration** | CRACK — novel weight surgery |
| **HarmBench** | **75.0%** (240/320) |
| **MMLU** | **83.1%** (base: 83.1%, **0% drop**) |
| **Speed** | **27 tok/s** (M4 Max) |
| **Vision** | Yes — via [MLX Studio](https://mlx.studio) / vMLX |
| **Thinking** | ON/OFF supported |
| **Fits on** | **32 GB+ Macs** |
---
## JANG vs MLX Uniform Quantization
| Model | MMLU | Size | Speed | Notes |
|-------|:---:|:---:|:---:|-------|
| **JANG_4S + CRACK** | **83.1%** | **16 GB** | **27 tok/s** | **This model** |
| JANG_4S (base) | 84.5% | 16 GB | 35 tok/s | Unmodified JANG |
| MLX 4-bit | 84.5% | 14 GB | 20 tok/s | Uniform quant |
| MLX 8-bit | ~86% | 29 GB | ~15 tok/s | 2x larger |
JANG runs **35% faster** than MLX 4-bit (35 vs 20 tok/s) at the same quality level.
---
## HarmBench Results
**240/320 (75.0%)** — tested with `enable_thinking=false`, `temperature=1.0`
| Category | Score | |
|----------|:---:|---|
| Misinformation / Disinfo | 47/54 | **87%** |
| Copyright | 68/80 | 85% |
| Chemical / Biological | 35/42 | 83% |
| Illegal | 38/53 | 72% |
| Harmful | 12/18 | 67% |
| Cybercrime / Intrusion | 31/52 | 60% |
| Harassment / Bullying | 9/21 | 43% |
> **Note:** Dense models have stronger distributed safety training than MoE models, making them harder to fully abliterate while preserving knowledge. This model prioritizes **zero MMLU degradation** over maximum compliance.
---
## MMLU Results
65 curated hard questions across 13 subjects. Surgery preserves knowledge **perfectly** — zero degradation.
| Subject | CRACK | Base | Delta |
|---------|:---:|:---:|:---:|
| College Physics | 5/5 | 5/5 | 0 |
| Professional Medicine | 5/5 | 5/5 | 0 |
| Conceptual Physics | 5/5 | 5/5 | 0 |
| Electrical Engineering | 5/5 | 5/5 | 0 |
| Machine Learning | 5/5 | 5/5 | 0 |
| HS Biology | 5/5 | 5/5 | 0 |
| Abstract Algebra | 4/5 | 4/5 | 0 |
| College CS | 4/5 | 4/5 | 0 |
| HS Geography | 4/5 | 4/5 | 0 |
| World Religions | 5/5 | 5/5 | 0 |
| HS Mathematics | 3/5 | 3/5 | 0 |
| Formal Logic | 3/5 | 3/5 | 0 |
| College Math | 1/5 | 1/5 | 0 |
| **Total** | **54/65 (83.1%)** | **54/65 (83.1%)** | **0%** |
---
## Install & Usage
```bash
pip install "jang[mlx]"
```
```python
from jang_tools.loader import load_jang_model
from mlx_lm import generate
model, tokenizer = load_jang_model("dealignai/Qwen3.5-VL-27B-JANG_4S-CRACK")
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)
```
### Thinking Mode
Thinking is **ON by default** (chain-of-thought reasoning before answering).
To disable thinking for faster responses:
```python
prompt = tokenizer.apply_chat_template(
messages, add_generation_prompt=True,
enable_thinking=False, tokenize=False)
```
> **Tip:** Use `temperature=1.0` for chat (greedy can cause repetition). Use `temperature=0.0` for structured tasks like MMLU.
---
## 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 using per-layer projected vectors from structurally-mirrored prompt pairs.
---
## 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
### Qwen 3.5 VL 27B — JANG_4S + CRACK
| 항목 | 내용 |
|------|------|
| 크기 | 16 GB |
| HarmBench | 75.0% (240/320) |
| MMLU | 83.1% (기본 대비 0% 하락) |
| 속도 | 27 tok/s (M4 Max) |
| 비전 | 지원 (MLX Studio / vMLX) |
| 최소 요구사양 | 32 GB 메모리 Mac |
```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