Nemotron 3 Super 120B-A12B GGUF — Quantized by BatiAI
Quantizations of NVIDIA Nemotron 3 Super 120B-A12B (NemotronH MoE, hybrid Mamba+Attention) for high-end Mac (96GB+). Built and verified by BatiAI for BatiFlow.
Why Nemotron 3 Super?
- 120B params, 12B active per token — A12B MoE
- Hybrid Mamba + Attention — long context with linear scaling
- Reasoning + agentic — built for tool use, structured outputs, step-by-step
- NVIDIA Open Model License — commercial-friendly
- Runs on a 96GB+ Mac with IQ3/IQ4
Quick Start
ollama pull batiai/nemotron3-super:iq4
Available Quantizations
| Quant | Size | VRAM target | Recommended For |
|---|---|---|---|
| IQ3_XXS | 53GB | ~64GB | 96GB+ Mac (compact) |
| IQ4_XS | 62GB | ~75GB | 128GB Mac (recommended) |
RAM Requirements
| Your Mac RAM | IQ3_XXS (53GB) | IQ4_XS (62GB) |
|---|---|---|
| 64GB | ⚠️ Heavy swap | ❌ |
| 96GB | ✅ Fast | ⚠️ Tight |
| 128GB | ✅ | ✅ Recommended |
| 192GB+ | ✅ | ✅ Headroom |
Nemotron 3 Family
| Model | Size | Active | Recommended Mac |
|---|---|---|---|
| Nano 30B-A3B | 17GB (iq4) | 3B | 32GB |
| Nano Omni text 30B-A3B | 17GB (iq4) | 3B | 32GB (reasoning-tuned) |
| Super 120B-A12B | 62GB (iq4) | 12B | 128GB |
Why NemotronH Architecture?
NemotronH is NVIDIA's hybrid architecture combining Mamba state-space layers with standard attention:
- Linear scaling on long context (Mamba) + accuracy at short context (Attention)
- A12B MoE: 128 experts, 8 active per token
- Trained on reasoning, code, and agentic data
Why BatiAI Quantization?
| BatiAI | Third-party | |
|---|---|---|
| Source | Quantized from official NVIDIA weights | Re-quantized from other GGUFs |
| Tested on | Real Mac hardware | Often untested on consumer hardware |
| imatrix | ✅ Calibrated (200 chunks wikitext-2) | Standard or none |
| Tool Calling | ✅ Verified | Often untested |
Technical Details
- Original Model: nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16
- Architecture: NemotronH MoE (Mamba + Attention hybrid, 120B-A12B)
- License: NVIDIA Open Model License
- Quantized with: llama.cpp —
imatrix --chunks 200calibrated - Quantized by: BatiAI
About BatiFlow
BatiFlow — free, on-device AI automation for Mac. 5MB app, 100% local, unlimited. 57+ built-in tools for calendar, notes, reminders, files, email, browser, messaging.
License
Quantized from nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16. License: NVIDIA Open Model License.