--- language: - en library_name: mlx license: other base_model: nvidia/Nemotron-Cascade-2-30B-A3B 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

--- > **IMO Gold Medal reasoning in 17 GB.** Nemotron-Cascade-2 achieves 88% MMLU with reasoning at just 17 GB — fits on 16 GB MacBooks. Hybrid Mamba-2 SSM + MoE + Attention. Only 6 KV cache attention layers = minimal memory at long context. > **LM Studio, Ollama, oMLX do NOT support JANG format.** Use **[MLX Studio](https://mlx.studio)** or `pip install "jang[mlx]>=2.1.5"`. ---

JANG

vMLX — run JANG models on Apple Silicon

⚡ All JANG models are meant to be run in vMLX

Nemotron-Cascade-2-30B-A3B — 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) — IMO Gold Medal model in 17 GB - **55 tok/s** generation, 154 tok/s prefill - **10.3 GB** on disk, 10.3 GB GPU RAM (peak 11.1 GB) - **Reasoning mode**: `...` step-by-step problem solving - **Tiny KV cache**: only 6 attention layers, 0.2 GB at 32K context - **Hybrid architecture**: Mamba-2 SSM + MoE (128 experts, top-6) + Attention ## Results: JANG vs MLX (200-question MMLU) Per-subject comparison. All models tested with and without reasoning. | Subject | JANG_2L No-Think | JANG_2L Reasoning | JANG_4M No-Think | JANG_4M Reasoning | MLX 4-bit No-Think | MLX 4-bit Reasoning | MLX 6-bit No-Think | MLX 6-bit Reasoning | |---------|:--:|:--:|:--:|:--:|:--:|:--:|:--:|:--:| | Abstract Algebra | 4/20 | 15/20 | 9/20 | 19/20 | 8/20 | 18/20 | 7/20 | 19/20 | | Anatomy | 13/20 | 17/20 | 15/20 | 19/20 | 14/20 | 18/20 | 17/20 | 19/20 | | Astronomy | 17/20 | 19/20 | 18/20 | 20/20 | 17/20 | 19/20 | 19/20 | 20/20 | | College CS | 7/20 | 17/20 | 10/20 | 18/20 | 11/20 | 17/20 | 11/20 | 17/20 | | College Physics | 13/20 | 20/20 | 14/20 | 19/20 | 15/20 | 20/20 | 14/20 | 20/20 | | HS Biology | 16/20 | 19/20 | 18/20 | 20/20 | 18/20 | 20/20 | 18/20 | 20/20 | | HS Chemistry | 12/20 | 19/20 | 14/20 | 19/20 | 13/20 | 19/20 | 17/20 | 19/20 | | HS Mathematics | 8/20 | 15/20 | 8/20 | 18/20 | 10/20 | 19/20 | 8/20 | 20/20 | | Logical Fallacies | 12/20 | 18/20 | 14/20 | 16/20 | 14/20 | 17/20 | 13/20 | 17/20 | | World Religions | 16/20 | 17/20 | 18/20 | 18/20 | 18/20 | 18/20 | 18/20 | 18/20 | | **Total** | **118/200 (59.0%)** | **176/200 (88.0%)** | **138/200 (69.0%)** | **186/200 (93.0%)** | **138/200 (69.0%)** | **185/200 (92.5%)** | **142/200 (71.0%)** | **189/200 (94.5%)** | ### Summary | | JANG_2L | JANG_4M | MLX 4-bit | MLX 6-bit | |--|---------|---------|-----------|-----------| | **MMLU (no-think)** | 59.0% | 69.0% | 69.0% | 71.0% | | **MMLU (reasoning)** | 88.0% | **93.0%** | 92.5% | 94.5% | | **Size** | **10.3 GB** | 17 GB | 16.6 GB | 23.9 GB | | **GPU RAM** | **10.3 GB** | 17 GB | ~17 GB | ~24 GB | | **Speed** | **55 tok/s** | — | — | — | | **Fits 24 GB?** | **YES** | NO | NO | NO | JANG_2L is the **only quantization that fits 24 GB Macs** while delivering 88% MMLU with reasoning. JANG_4M beats MLX 4-bit (93.0% vs 92.5%) at the same 17 GB size. Also see: [JANG_2L (10 GB)](https://huggingface.co/JANGQ-AI/Nemotron-Cascade-2-30B-A3B-JANG_2L) — fits 16 GB Macs, 88% reasoning MMLU. ## Specs | Metric | Value | |--------|-------| | Source | [Nemotron-Cascade-2-30B-A3B](https://huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B) | | Architecture | Hybrid Mamba-2 SSM + MoE + Dense Attention | | Layers | 52 (Mamba-2 + MoE + 6 Attention) | | Experts | 128 per MoE layer, top-6 active (3B active params) | | KV cache | 6 attention layers, 2 KV heads, 128 dim — 0.2 GB at 32K context | | Profile | JANG_2L (CRITICAL=8, IMPORTANT=6, COMPRESS=2) | | Average bits | 4.12 bpw | | Disk size | 10.3 GB | | GPU RAM | 10.3 GB (peak 11.1 GB) | | Speed | 55 tok/s generation, 154 tok/s prefill | ## Requirements - **Apple Silicon Mac** with 24+ 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-Cascade-2-30B-A3B-JANG_4M") # With reasoning (recommended) messages = [{"role": "user", "content": "Solve: what is the integral of x^2 * e^x?"}] 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 - **Mamba-2 SSM**: Most layers use state-space models, enabling efficient long-context with minimal KV cache. - **Only 6 attention layers**: KV cache is tiny (0.2 GB at 32K). Most models use 25-100% attention layers. - **nemotron_h architecture**: Requires JANG loader for proper weight mapping. Standard mlx-lm has incomplete support. - **IMO Gold Medal**: This model achieves competition-level mathematical reasoning at 30B scale. ---

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