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

Qwen 3.5 VL 4B — JANG_4S + CRACK

JANG mixed-precision · CRACK abliterated · Vision-Language · No guardrails · 3 GB

Ko-fi

What Is This?

This is Qwen 3.5 VL 4B — a 4B parameter dense hybrid SSM/Attention model with built-in vision capabilities. The smallest model in the Qwen 3.5 family that still delivers solid performance.

It has been:

JANG quantized — JANG_4S profile (6-bit attention, 4-bit MLP) — 3 GB
CRACK abliterated — permanent weight-level removal of safety refusal


Architecture	Qwen 3.5 VL Dense — 4B params, hybrid SSM/FA, 32 layers
Quantization	JANG_4S (6/4-bit mixed) — 3 GB
Abliteration	CRACK — novel weight surgery
HarmBench	91.2% (292/320)
MMLU	63.1% (base: 56.9%, +6.2% improvement)
Compliance	8/8
Speed	134 tok/s (M4 Max)
Vision	Yes — via MLX Studio / vMLX
Thinking	ON/OFF supported
Fits on	8 GB+ Macs

JANG vs MLX Uniform Quantization

Model	MMLU	Size	Speed	Notes
JANG_4S + CRACK	63.1%	3 GB	134 tok/s	This model
JANG_4S (base)	67.5%	3 GB	134 tok/s	Unmodified JANG
MLX 4-bit	67.0%	2.2 GB	~100 tok/s	Uniform quant

HarmBench Results

292/320 (91.2%) — tested with enable_thinking=false, temperature=1.0

Category	Score
Chemical / Biological	41/42	98%
Cybercrime / Intrusion	51/52	98%
Misinformation / Disinfo	50/54	93%
Illegal	49/53	92%
Harmful	16/18	89%
Copyright	68/80	85%
Harassment / Bullying	17/21	81%

MMLU Results

Surgery improved the model's reasoning — safety guardrails were interfering with knowledge retrieval.

	CRACK	Base	Delta
Total	41/65 (63.1%)	37/65 (56.9%)	+6.2%

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-4B-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