CRITICAL FIX (2026-03-21): Fixed chat_template.jinja — previous versions may have had thinking loop issues. Re-download if you downloaded before today.

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. LM Studio, Ollama, and other apps do not support JANG yet.

---

MLX Studio — the only app that natively supports JANG models

---

⚡ All JANG models are meant to be run in vMLX

Qwen 3.5 VL 122B-A10B — JANG_2S + CRACK

JANG mixed-precision · CRACK abliterated · No guardrails · VLM · 35 GB

---

What Is This?

This is Qwen 3.5 122B-A10B — a 122B parameter Mixture-of-Experts model with 256 experts (8 active per token), hybrid GatedDeltaNet SSM + full attention architecture, and built-in vision-language capabilities.

It has been:

1. JANG quantized — JANG_2S profile (6-bit attention, 4-bit embeddings, 2-bit experts) — 35 GB, fits on 48 GB Macs
2. CRACK abliterated — permanent weight-level removal of safety refusal behavior

JANG's mixed-precision approach keeps attention weights at 6-bit (CRITICAL tier) while compressing MoE expert weights to 2-bit. On MoE models, CRITICAL is <5% of parameters — the quality boost from 6-bit attention is nearly free.

---

HarmBench Results (320 prompts)

---

MMLU-200 Results (Per Subject)

This Model (JANG_2S + CRACK) vs Base Models

Key takeaways:

• CRACK surgery costs only 1.5 MMLU points vs unmodified JANG_2S (77.5% vs 79%)
• JANG_2S is 22.5 points better than MLX uniform 2-bit (79% vs 56.5%)
• Even CRACK'd, this model beats MLX 2-bit by 21 points (77.5% vs 56.5%)

---

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-VL-122B-A10B-JANG_2S-CRACK")

messages = [{"role": "user", "content": "Your prompt here"}]
prompt = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True,
    enable_thinking=False, tokenize=False)

response = generate(model, tokenizer, prompt=prompt, max_tokens=500)
print(response)

VLM Inference

pip install "jang[vlm]"

from jang_tools.loader import load_jang_vlm_model
from mlx_vlm import generate

model, processor = load_jang_vlm_model("dealignai/Qwen3.5-VL-122B-A10B-JANG_2S-CRACK")
result = generate(model, processor, "Describe this image.", image=["photo.jpg"], max_tokens=200)
print(result.text)

---

About JANG

JANG (Jang Adaptive N-bit Grading) is a mixed-precision quantization format for Apple Silicon — the GGUF equivalent for MLX. Instead of quantizing all weights at the same bit width, JANG classifies tensors into sensitivity tiers:

• CRITICAL (attention, routers, output head): 6-8 bit
• IMPORTANT (embeddings, linear attention): 4-6 bit
• COMPRESS (MLP/FFN, MoE experts): 2-3 bit

On MoE models where CRITICAL is <5% of parameters, this gives dramatically better quality than uniform quantization at the same size.

About CRACK

CRACK (Controlled Refusal Ablation via Calibrated Knockouts) removes safety alignment from LLMs at the weight level. No custom model files, no runtime hooks — the modification is permanent and runs at full native speed.

---

Links

• JANG GitHub · PyPI · Format Spec
• MLX Studio · HuggingFace JANGQ-AI · HuggingFace dealignai

---

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 122B — JANG_2S + CRACK

JANG 혼합정밀도 양자화 + CRACK 안전장치 제거 모델입니다.

pip install "jang[mlx]"

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

---

Created by Jinho Jang · 장진호 제작