CoolFace
Modelpublic

JANGQ-AI/Nemotron-3-Super-120B-A12B-JANG_4M

sourceHugging Faceotherupdated 20d agoView on Hugging Face
0likes175downloads
Model Card

<p align="center"> <a href="https://mlx.studio"><img src="https://raw.githubusercontent.com/jjang-ai/jangq/main/assets/mlx-studio-light.png" alt="MLX Studio" width="500"></a> </p>

<p align="center"> <a href="https://mlx.studio"><img src="https://mlx.studio/assets/screenshots/mlx-studio-featured.png?v=1" alt="MLX Studio App" width="600"></a> </p>

<h4 align="center"><a href="https://mlx.studio">MLX Studio</a> — the only app that natively supports JANG models with reasoning</h4>

93% MMLU at same size as MLX 4-bit. JANG_4M matches MLX 4-bit quality with 8-bit attention protection. Hybrid Mamba-2 SSM + Latent MoE + Attention.
LM Studio, Ollama, oMLX do NOT support JANG format. Use [MLX Studio](https://mlx.studio) or \pip install "jang[mlx]>=2.1.5"\.

<p align="center"> <img src="https://raw.githubusercontent.com/jjang-ai/jangq/main/assets/jangq-logo-dark.png" alt="JANG" width="300"> </p>

<p align="center"> <a href="https://vmlx.net"><img src="vmlx-app.png" alt="vMLX — run JANG models on Apple Silicon" width="820"></a> </p>

<h3 align="center">⚡ All JANG models are meant to be run in <a href="https://vmlx.net">vMLX</a></h3>

<h3 align="center">Nemotron-3-Super-120B-A12B — JANG_4M (4.1-bit, 8-bit attention) — Reasoning</h3> <p align="center"><b>JANG</b> — Jang Adaptive N-bit Grading | The GGUF Equivalent for MLX</p>

<p align="center"> <a href="https://github.com/jjang-ai/jangq"><img src="https://img.shields.io/badge/GitHub-Source_Code-blue?logo=github" alt="GitHub"></a>&nbsp; <a href="https://pypi.org/project/jang/"><img src="https://img.shields.io/pypi/v/jang?label=PyPI&color=green" alt="PyPI"></a>&nbsp; <a href="https://jangq.ai"><img src="https://img.shields.io/badge/Web-jangq.ai-orange" alt="Website"></a>&nbsp; <a href="https://x.com/dealignai"><img src="https://img.shields.io/badge/X-@dealignai-black?logo=x" alt="X/Twitter"></a> </p>

JANG is fully open-source. Quantization engine, research, and full commit history: github.com/jjang-ai/jangq. Created by Jinho Jang.

Key Features

  • —93.0% MMLU (200 questions, reasoning mode) — matches MLX 4-bit at same size
  • —55.1 tok/s generation, 154 tok/s prefill
  • —63 GB on disk, 61.2 GB GPU RAM
  • —Reasoning mode: \<think>...</think>\ step-by-step problem solving
  • —Hybrid architecture: 40 Mamba-2 SSM + 40 Latent MoE (512 experts) + 8 Dense Attention layers
  • —bfloat16 compute: auto-detected for 512-expert models

Results: JANG vs MLX (200-question MMLU)

Per-subject comparison. All models tested with and without reasoning using identical methodology.

SubjectJANG_4M No-ThinkJANG_4M ReasoningJANG_2L No-ThinkJANG_2L ReasoningMLX 4-bit No-ThinkMLX 4-bit Reasoning
Abstract Algebra10/2019/2012/2016/209/2019/20
Anatomy15/2018/2015/2017/2014/2018/20
Astronomy19/2019/2019/2019/2019/2019/20
College CS13/2017/2013/2015/2014/2017/20
College Physics14/2019/2014/2018/2013/2020/20
HS Biology19/2020/2019/2018/2018/2020/20
HS Chemistry15/2018/2015/2016/2016/2019/20
HS Mathematics6/2018/208/2018/206/2018/20
Logical Fallacies17/2019/2017/2018/2017/2018/20
World Religions17/2019/2018/2017/2016/2019/20
Total145/200 (72.5%)186/200 (93.0%)150/200 (75.0%)172/200 (86.0%)142/200 (71.0%)187/200 (93.5%)

Summary

JANG_4MJANG_2LMLX 4-bitMLX 3-bit
MMLU (no-think)72.5%75.0%71.0%Crashes
MMLU (reasoning)93.0%86.0%93.5%Crashes
Size63 GB43 GB63 GBN/A
GPU RAM61.2 GB42.4 GB63.3 GBN/A
Speed55.1 tok/s51.6 tok/s59.8 tok/sN/A
Fits 64 GB?YESYESYESN/A

JANG4M nearly ties MLX 4-bit (93.0% vs 93.5%) at the same 63 GB size with 8-bit attention protection. MLX 3-bit cannot be created — \`mlxlm.convert\` crashes on Nemotron's mtp.* weights. Only JANG can produce sub-4-bit quantizations.

Also see: JANG_2L (43 GB) — 20 GB smaller, fits 64 GB Macs, 75% no-think / 86% reasoning.

Specs

MetricValue
SourceNVIDIA-Nemotron-3-Super-120B-A12B-FP8
ArchitectureHybrid Mamba-2 SSM + Latent MoE + Dense Attention
Layers88 (40 Mamba-2 + 40 MoE + 8 Attention)
Experts512 per MoE layer, top-22 active (12B active params)
ProfileJANG_4M (CRITICAL=8, IMPORTANT=4, COMPRESS=4)
Average bits4.10 bpw
Disk size63 GB
GPU RAM61.2 GB (peak 66 GB)
Speed55.1 tok/s generation, 154 tok/s prefill
Computebfloat16 (auto-detected)

Requirements

  • —Apple Silicon Mac with 64+ GB unified memory
  • —[MLX Studio](https://mlx.studio) or \pip install "jang[mlx]>=2.1.5"\

Quick Start

\\\bash pip install "jang[mlx]>=2.1.5" \\\

\\\`python from jangtools.loader import loadjangmodel from mlxlm import generate

model, tokenizer = loadjangmodel("JANGQ-AI/Nemotron-3-Super-120B-A12B-JANG_4M")

With reasoning

messages = [{"role": "user", "content": "Explain quantum computing."}] prompt = tokenizer.applychattemplate(messages, tokenize=False, addgenerationprompt=True, enablethinking=True) result = generate(model, tokenizer, prompt=prompt, maxtokens=2048)

Without reasoning (faster)

prompt = tokenizer.applychattemplate(messages, tokenize=False, addgenerationprompt=True, enablethinking=False) result = generate(model, tokenizer, prompt=prompt, maxtokens=100) \\\`

Technical Notes

  • —Latent MoE: Nemotron-H compresses hidden states 4096→1024 before expert routing. JANG loader handles this automatically.
  • —bfloat16: Auto-detected for 512-expert models. Prevents float16 overflow. Zero quality impact.
  • —trust_remote_code: Custom Python files included (modelingnemotronh.py, configurationnemotronh.py).

<p align="center"> <b>JANG</b> — Created by <a href="https://jangq.ai">Jinho Jang</a> (eric@jangq.ai) · <a href="https://x.com/dealignai">@dealignai</a><br> <a href="https://github.com/jjang-ai/jangq">GitHub</a> · <a href="https://pypi.org/project/jang/">PyPI</a> · <a href="https://huggingface.co/JANGQ-AI">HuggingFace</a> </p>

한국어

Nemotron-3-Super-120B JANG_4M — MLX 4-bit과 동일한 크기(63 GB)에서 93% MMLU 달성.

JANG_4MJANG_2LMLX 4-bit
MMLU (추론 없음)72.5%75.0%71.0%
MMLU (추론 포함)93.0%86.0%93.5%
크기63 GB43 GB63 GB
속도55.1 tok/s51.6 tok/s59.8 tok/s

\\\bash pip install "jang[mlx]>=2.1.5" \\\