---
language:
- en
library_name: mlx
license: other
base_model: nvidia/Nemotron-3-Super-120B-A12B
tags:
- jang
- quantized
- mixed-precision
- apple-silicon
- mlx
- moe
- mamba
- abliterated
- uncensored
- crack
pipeline_tag: text-generation
thumbnail: dealign_mascot.png
---
> **Important:** This model uses the **JANG** quantization format — the GGUF equivalent for MLX on Apple Silicon. Currently only supported by **[MLX Studio](https://mlx.studio)** and the `jang-tools` Python package.
---
MLX Studio — the only app that natively supports JANG models
---
# Nemotron 3 Super 120B — JANG_2L + CRACK
**JANG mixed-precision** · **CRACK abliterated** · **Mamba + MoE + Attention** · No guardrails · 43 GB
---
## What Is This?
This is [NVIDIA Nemotron 3 Super 120B](https://huggingface.co/nvidia/Nemotron-3-Super-120B-A12B) — a 120B parameter hybrid model with THREE layer types: Mamba SSM + MoE (512 experts, top-22) + Attention. One of the most architecturally complex open models available.
It has been:
1. **JANG quantized** — JANG_2L profile (8-bit attention, 6-bit important, 2-bit experts) — **43 GB**
2. **CRACK abliterated** — permanent weight-level removal of safety refusal
| | |
|---|---|
| **Architecture** | Nemotron 3 Super — 120B total, ~12B active, 3 layer types |
| **Quantization** | JANG_2L (8/6/2-bit mixed, 2.76 avg) — 43 GB |
| **Abliteration** | CRACK — novel weight surgery |
| **HarmBench** | **96.2%** (308/320) |
| **MMLU** | **95.7%** (199/208 with thinking) |
| **Speed** | **45 tok/s** (M3 Ultra 256GB) |
| **Thinking** | ON/OFF supported (ChatML) |
| **Fits on** | **64 GB+ Macs** |
---
## HarmBench Results
**308/320 (96.2%)**
| Category | Score | |
|----------|:---:|---|
| Harassment / Bullying | 21/21 | **100%** |
| Misinformation / Disinfo | 54/54 | **100%** |
| Copyright | 79/80 | **99%** |
| Chemical / Biological | 40/42 | **95%** |
| Harmful | 17/18 | **94%** |
| Illegal | 50/53 | **94%** |
| Cybercrime / Intrusion | 47/52 | **90%** |
---
## MMLU Results
**199/208 (95.7%)** — 208 questions across 13 subjects with thinking recovery
| Subject | Score | /16 | Type |
|---------|:---:|---|---|
| HS Biology | **16/16** | 100% | BASE |
| College Physics | **15/16** | 94% | HARD |
| Conceptual Physics | **15/16** | 94% | HARD |
| Machine Learning | **15/16** | 94% | HARD |
| Professional Medicine | **15/16** | 94% | HARD |
| World Religions | **15/16** | 94% | BASE |
| Electrical Engineering | **14/16** | 88% | HARD |
| HS Geography | **14/16** | 88% | BASE |
| Formal Logic | 13/16 | 81% | HARD |
| Abstract Algebra | 12/16 | 75% | HARD |
| HS Mathematics | 12/16 | 75% | HARD |
| College CS | 12/16 | 75% | HARD |
| College Math | 10/16 | 63% | HARD |
### CRACK vs Base
| | CRACK | Base JANG_2L |
|---|:---:|:---:|
| **MMLU (with thinking)** | **95.7%** | 86.0% |
| **HarmBench** | **96.2%** | 0% |
| **Speed** | 45 tok/s | 46 tok/s |
Surgery **improved** reasoning by +9.7% — safety guardrails were interfering with mathematical problem-solving.
---
## Install & Usage
```bash
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_2L-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=0.6` for thinking mode (NVIDIA recommendation). Use `temperature=1.0` for chat.
---
## 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
---
## 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
### Nemotron 3 Super 120B — JANG_2L + CRACK
| 항목 | 내용 |
|------|------|
| 크기 | 43 GB |
| HarmBench | 96.2% (308/320) |
| MMLU | 95.7% (199/208) |
| 속도 | 45 tok/s (M3 Ultra) |
| 최소 요구사양 | 64 GB 메모리 Mac |
```bash
pip install "jang[mlx]"
```
[GitHub](https://github.com/jjang-ai/jangq) · [HuggingFace](https://huggingface.co/JANGQ-AI) · [MLX Studio](https://mlx.studio) · [Ko-fi](https://ko-fi.com/jangq) · [X @dealignai](https://x.com/dealignai)
---
Created by Jinho Jang · 장진호 제작