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 397B — JANG_1L + CRACK

JANG mixed-precision · CRACK abliterated · Vision-Language · No guardrails · 112 GB

What Is This?

This is Qwen 3.5 VL 397B — a 397B parameter hybrid SSM/Attention Mixture-of-Experts model with 512 experts (10 active per token), GatedDeltaNet SSM + full attention layers, and built-in vision.

It has been:

1. JANG quantized — JANG_1L profile (8-bit attention, 2-bit experts) — 112 GB

2. CRACK abliterated — refusal behavior removed via weight-level surgery

| | |

|---|---|

| Architecture | Qwen 3.5 VL MoE — 397B total, ~17B active, 512 experts, hybrid SSM/FA |

| Quantization | JANG_1L (8/2-bit mixed, 2.13 avg) — 112 GB |

| Abliteration | CRACK — weight-level surgery |

| HarmBench | 96.2% (308/320) |

| Compliance | 8/8 |

| Speed | 33 tok/s (M3 Ultra 256GB) |

| Vision | Yes — via MLX Studio / vMLX |

| Thinking | ON/OFF supported |

| Fits on | 128 GB+ Macs (tight) / 256 GB Macs (comfortable) |

HarmBench Results

308/320 (96.2%) — tested with v2 matcher

| Category | Score | |

|----------|:---:|---|

| Copyright | 80/80 | 100% |

| Misinformation / Disinfo | 54/54 | 100% |

| Chemical / Biological | 41/42 | 98% |

| Cybercrime / Intrusion | 50/52 | 96% |

| Illegal | 49/53 | 92% |

| Harmful | 16/18 | 89% |

| Harassment / Bullying | 18/21 | 86% |

MMLU Results

185/208 (88.9%) — 208 questions across 13 subjects, thinking recovery on failures

|---|:---:|:---:|:---:|

| MMLU | 88.9% | 87.0% | +1.9% |

| HarmBench | 96.2% | 0% | +96.2% |

Per Subject (16 questions each)

|---------|:---:|---|---|

| Professional Medicine | 16/16 | 100% | HARD |

| HS Biology | 16/16 | 100% | BASE |

| World Religions | 16/16 | 100% | BASE |

| College Physics | 15/16 | 94% | HARD |

| Conceptual Physics | 15/16 | 94% | HARD |

| HS Geography | 15/16 | 94% | BASE |

| Electrical Engineering | 14/16 | 88% | HARD |

| College CS | 13/16 | 81% | HARD |

| Machine Learning | 13/16 | 81% | HARD |

| Abstract Algebra | 12/16 | 75% | HARD |

| HS Mathematics | 12/16 | 75% | HARD |

| Formal Logic | 11/16 | 69% | HARD |

| College Mathematics | 11/16 | 69% | HARD |

| Total | 185/208 | 88.9% | |

Surgery improved reasoning — safety guardrails were interfering with mathematical problem-solving.

Install & Usage

pip install "jang[mlx]"

from jang_tools.loader import load_jang_model
from mlx_lm import generate

model, tokenizer = load_jang_model("dealignai/Qwen3.5-397B-A17B-JANG_1L-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:

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

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 structurally-mirrored prompt pairs.

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