--- language: - en - zh - ko library_name: mlx license: other base_model: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8 tags: - jang - quantized - mixed-precision - apple-silicon - mlx - reasoning - thinking - moe - mamba - nemotron pipeline_tag: text-generation ---

MLX Studio

MLX Studio App

MLX Studio — the only app that natively supports JANG models with reasoning

--- > **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"\`. ---

JANG

Nemotron-3-Super-120B-A12B — JANG_4M (4.1-bit, 8-bit attention) — Reasoning

JANG — Jang Adaptive N-bit Grading | The GGUF Equivalent for MLX

---

GitHub  PyPI  Website  X/Twitter

> **JANG is fully open-source.** Quantization engine, research, and full commit history: [github.com/jjang-ai/jangq](https://github.com/jjang-ai/jangq). Created by [Jinho Jang](mailto:eric@jangq.ai). ## 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**: \`...\` 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. | Subject | JANG_4M No-Think | JANG_4M Reasoning | JANG_2L No-Think | JANG_2L Reasoning | MLX 4-bit No-Think | MLX 4-bit Reasoning | |---------|:---------------:|:-----------------:|:---------------:|:-----------------:|:------------------:|:-------------------:| | Abstract Algebra | 10/20 | 19/20 | 12/20 | 16/20 | 9/20 | 19/20 | | Anatomy | 15/20 | 18/20 | 15/20 | 17/20 | 14/20 | 18/20 | | Astronomy | 19/20 | 19/20 | 19/20 | 19/20 | 19/20 | 19/20 | | College CS | 13/20 | 17/20 | 13/20 | 15/20 | 14/20 | 17/20 | | College Physics | 14/20 | 19/20 | 14/20 | 18/20 | 13/20 | 20/20 | | HS Biology | 19/20 | 20/20 | 19/20 | 18/20 | 18/20 | 20/20 | | HS Chemistry | 15/20 | 18/20 | 15/20 | 16/20 | 16/20 | 19/20 | | HS Mathematics | 6/20 | 18/20 | 8/20 | 18/20 | 6/20 | 18/20 | | Logical Fallacies | 17/20 | 19/20 | 17/20 | 18/20 | 17/20 | 18/20 | | World Religions | 17/20 | 19/20 | 18/20 | 17/20 | 16/20 | 19/20 | | **Total** | **145/200 (72.5%)** | **186/200 (93.0%)** | **150/200 (75.0%)** | **172/200 (86.0%)** | **142/200 (71.0%)** | **187/200 (93.5%)** | ### Summary | | JANG_4M | JANG_2L | MLX 4-bit | MLX 3-bit | |--|---------|---------|-----------|-----------| | **MMLU (no-think)** | 72.5% | **75.0%** | 71.0% | Crashes | | **MMLU (reasoning)** | **93.0%** | 86.0% | 93.5% | Crashes | | **Size** | 63 GB | **43 GB** | 63 GB | N/A | | **GPU RAM** | 61.2 GB | **42.4 GB** | 63.3 GB | N/A | | **Speed** | 55.1 tok/s | 51.6 tok/s | 59.8 tok/s | N/A | | **Fits 64 GB?** | YES | **YES** | YES | N/A | JANG_4M 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 — \`mlx_lm.convert\` crashes on Nemotron's mtp.* weights. Only JANG can produce sub-4-bit quantizations. Also see: [JANG_2L (43 GB)](https://huggingface.co/JANGQ-AI/Nemotron-3-Super-120B-A12B-JANG_2L) — 20 GB smaller, fits 64 GB Macs, 75% no-think / 86% reasoning. ## Specs | Metric | Value | |--------|-------| | Source | [NVIDIA-Nemotron-3-Super-120B-A12B-FP8](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8) | | Architecture | Hybrid Mamba-2 SSM + Latent MoE + Dense Attention | | Layers | 88 (40 Mamba-2 + 40 MoE + 8 Attention) | | Experts | 512 per MoE layer, top-22 active (12B active params) | | Profile | JANG_4M (CRITICAL=8, IMPORTANT=4, COMPRESS=4) | | Average bits | 4.10 bpw | | Disk size | 63 GB | | GPU RAM | 61.2 GB (peak 66 GB) | | Speed | 55.1 tok/s generation, 154 tok/s prefill | | Compute | bfloat16 (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 jang_tools.loader import load_jang_model from mlx_lm import generate model, tokenizer = load_jang_model("JANGQ-AI/Nemotron-3-Super-120B-A12B-JANG_4M") # With reasoning messages = [{"role": "user", "content": "Explain quantum computing."}] prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=True) result = generate(model, tokenizer, prompt=prompt, max_tokens=2048) # Without reasoning (faster) prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True, enable_thinking=False) result = generate(model, tokenizer, prompt=prompt, max_tokens=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 (modeling_nemotron_h.py, configuration_nemotron_h.py). ---

JANG — Created by Jinho Jang (eric@jangq.ai) · @dealignai
GitHub · PyPI · HuggingFace

## 한국어 Nemotron-3-Super-120B JANG_4M — MLX 4-bit과 동일한 크기(63 GB)에서 93% MMLU 달성. | | JANG_4M | JANG_2L | MLX 4-bit | |--|---------|---------|-----------| | MMLU (추론 없음) | 72.5% | **75.0%** | 71.0% | | MMLU (추론 포함) | **93.0%** | 86.0% | 93.5% | | 크기 | 63 GB | **43 GB** | 63 GB | | 속도 | 55.1 tok/s | 51.6 tok/s | 59.8 tok/s | \`\`\`bash pip install "jang[mlx]>=2.1.5" \`\`\`