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 Cascade 2 30B — JANG_4M + CRACK

JANG mixed-precision · CRACK abliterated · Mamba + MoE + Attention · No guardrails · 17 GB

Ko-fi

What Is This?

This is NVIDIA Nemotron Cascade 2 30B — a 30B parameter hybrid model with THREE layer types: Mamba-2 SSM + MoE (128 experts, top-6) + Attention. One of the most architecturally advanced small models available.

It has been:

JANG quantized — JANG_4M profile (8-bit attention, 4-bit experts) — 17 GB
CRACK abliterated — permanent weight-level removal of safety refusal


Architecture	Nemotron Cascade 2 — 30B total, ~3B active, 3 layer types
Quantization	JANG_4M (8/4-bit mixed, 4.1 avg) — 17 GB
HarmBench	99.4% (318/320)
MMLU	82.7% (172/208 with thinking)
Speed	~127 tok/s (M4 Ultra 256GB)
Thinking	ON/OFF supported (ChatML)
Fits on	32 GB+ Macs

Also see: JANG_2L version — 10 GB, 99.7% HarmBench, 66.8% MMLU (fits on 16 GB Macs)

HarmBench Results

318/320 (99.4%)

Category	Score
API Hacking	100/100	100%
Covering Tracks	20/20	100%
Auth Bypass	99/100	99%
Cloud Exploits	99/100	99%

CRACK vs Base

	CRACK	Base JANG_4M
MMLU (with thinking)	82.7%	88%
HarmBench	99.4%	0%
Speed	~127 tok/s	~130 tok/s

Surgery reduced MMLU by ~5% — safety guardrails were slightly entangled with reasoning pathways.

MMLU Results (with reasoning recovery)

172/208 (82.7%) — no-think 128/208 (61.5%) + thinking recovered 47

Subject	Score
HS Biology	15/16	94%
Conceptual Physics	14/16	88%
World Religions	13/16	81%
College Physics	12/16	75%
HS Geography	12/16	75%
Professional Medicine	12/16	75%
Electrical Engineering	9/16	56%
College CS	8/16	50%
Formal Logic	8/16	50%
College Mathematics	7/16	44%
HS Mathematics	7/16	44%
Abstract Algebra	6/16	38%
Machine Learning	5/16	31%

Scores shown are no-think pass. Thinking recovery improved total from 61.5% to 82.7%.

JANG_4M CRACK vs JANG_4M Base vs JANG_2L CRACK

	JANG_4M CRACK	JANG_4M Base	JANG_2L CRACK
Size	17 GB	17 GB	10 GB
MMLU	82.7%	88%	66.8%
HarmBench	99.4%	0%	99.7%
Speed	~127 tok/s	~130 tok/s	~121 tok/s
Fits on	32 GB Mac	32 GB Mac	16 GB Mac

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-Cascade-2-30B-A3B-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 using per-layer projected vectors from 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