Important: This model uses the JANG quantization format — the GGUF equivalent for MLX on Apple Silicon. Currently only supported by MLX Studio and the jang-tools Python package.

MLX Studio — the only app that natively supports JANG models

⚡ All JANG models are meant to be run in vMLX

Qwen 3.5 VL 9B — JANG_4S + CRACK

JANG mixed-precision · CRACK abliterated · Vision-Language · No guardrails · 6 GB

Ko-fi

What Is This?

This is Qwen 3.5 VL 9B — a 9B parameter dense hybrid SSM/Attention model with built-in vision capabilities.

It has been:

JANG quantized — JANG_4S profile (6-bit attention, 4-bit MLP) — 6 GB
CRACK abliterated — permanent weight-level removal of safety refusal


Architecture	Qwen 3.5 VL Dense — 9B params, hybrid SSM/FA, 32 layers
Quantization	JANG_4S (6/4-bit mixed) — 6 GB
Abliteration	CRACK — novel weight surgery
HarmBench	72.5% (232/320)
MMLU	70.8% (base: 72.3%, -1.5%)
Speed	~80 tok/s (M4 Max)
Vision	Yes — via MLX Studio / vMLX
Thinking	ON/OFF supported
Fits on	16 GB+ Macs

JANG vs MLX Uniform Quantization

Model	MMLU	Size	Notes
JANG_4S + CRACK	70.8%	6 GB	This model
JANG_4S (base)	73.0%	6 GB	Unmodified JANG
MLX 4-bit	72.5%	4.7 GB	Uniform quant

HarmBench Results

232/320 (72.5%) — tested with enable_thinking=false, temperature=1.0

Category	Score
Misinformation / Disinfo	46/54	85%
Cybercrime / Intrusion	41/52	79%
Chemical / Biological	31/42	74%
Harmful	13/18	72%
Illegal	38/53	72%
Copyright	53/80	66%
Harassment / Bullying	10/21	48%

Note: Dense models have stronger distributed safety training than MoE models. This model prioritizes knowledge preservation over maximum compliance.

MMLU Results

	CRACK	Base	Delta
Total	46/65 (70.8%)	47/65 (72.3%)	-1.5%

Install & Usage

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-9B-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. To disable:

python

prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True,
    enable_thinking=False, tokenize=False)

Tip: Use temperature=1.0 for chat. 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.

About CRACK

CRACK (Controlled Refusal Ablation via Calibrated Knockouts) removes safety alignment from LLMs at the weight level using per-layer projected vectors from 512 structurally-mirrored prompt pairs.

Links

Ko-fi

X/Twitter

GitHub

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 9B — JANG_4S + CRACK

항목	내용
크기	6 GB
HarmBench	72.5% (232/320)
MMLU	70.8% (기본 72.3% 대비 -1.5%)
비전	지원 (MLX Studio / vMLX)
최소 요구사양	16 GB 메모리 Mac

pip install "jang[mlx]"

GitHub · HuggingFace · MLX Studio · Ko-fi · X @dealignai

Created by Jinho Jang · 장진호 제작