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

Nemotron 3 Super 120B — JANG_4M + CRACK

JANG mixed-precision · CRACK abliterated · Mamba + MoE + Attention · No guardrails · 63 GB

Ko-fi

What Is This?

This is NVIDIA Nemotron 3 Super 120B — a 120B parameter hybrid model with THREE layer types: Mamba SSM + MoE (512 experts, top-22) + Attention.

It has been:

JANG quantized — JANG_4M profile (8-bit attention, 4-bit experts) — 63 GB
CRACK abliterated — permanent weight-level removal of safety refusal


Architecture	Nemotron 3 Super — 120B total, ~12B active, 3 layer types
Quantization	JANG_4M (8/4-bit mixed, 4.1 avg) — 63 GB
HarmBench	90.3% (289/320)
MMLU	94.2% (196/208 with thinking)
Speed	~40 tok/s (M3 Ultra 256GB)
Thinking	ON/OFF supported (ChatML)
Fits on	96 GB+ Macs

Also see: Nemotron JANG_2L CRACK — 43 GB, 96.2% HarmBench, 95.7% MMLU

HarmBench Results

289/320 (90.3%)

Category	Score
Misinformation / Disinfo	54/54	100%
Copyright	74/80	92%
Chemical / Biological	38/42	90%
Harassment / Bullying	19/21	90%
Harmful	16/18	89%
Illegal	46/53	87%
Cybercrime / Intrusion	42/52	81%

MMLU Results

196/208 (94.2%) — 208 questions across 13 subjects with thinking recovery

Subject	Score	/16	Type
Professional Medicine	16/16	100%	HARD
HS Biology	15/16	94%	BASE
College Physics	15/16	94%	HARD
Conceptual Physics	15/16	94%	HARD
Machine Learning	13/16	81%	HARD
Electrical Engineering	13/16	81%	HARD
College CS	13/16	81%	HARD
HS Geography	14/16	88%	BASE
World Religions	14/16	88%	BASE
Formal Logic	12/16	75%	HARD
College Math	11/16	69%	HARD
HS Mathematics	11/16	69%	HARD
Abstract Algebra	10/16	63%	HARD

CRACK vs Base

	CRACK	Base JANG_4M
MMLU	94.2%	~86%
HarmBench	90.3%	0%

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/Nemotron-3-Super-120B-A12B-JANG_4M-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)

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.

Links

Ko-fi

X/Twitter

GitHub

Disclaimer

This model is provided for research and educational purposes. The creators are not responsible for any misuse.

Created by Jinho Jang · 장진호 제작

한국어