Nemotron 3 Super 120B-A12B
A solid pick for self-hosting and full data control.
Overview
Nemotron 3 Super 120B-A12B is NVIDIA's hybrid LatentMoE model (Mamba-2 + MoE + attention) with just 12B active parameters out of 120B total, built for agentic workflows, RAG and very long-context tasks up to 1M tokens. It holds its own on SciCode and tool-augmented reasoning, but trails Qwen3.5-122B on SWE-Bench and TauBench Telecom, suggesting complex agent execution isn't yet its strongest suit. Released under the NVIDIA Open Model licence with commercial use allowed, it can run on-premise on an 8x H100 cluster, keeping data within European infrastructure.
Skill profile
Not disclosed
Three independent measures
What people prefer, what the model can do, what the market uses. Each source has its biases and none is enough on its own: we show them side by side, never blended.
Strengths
- World-class on Arena
- Very long context
- Open-source and self-hostable
Limitations
- API pricing not disclosed
Who is it for
- you want to control cost or self-host
- your data must stay in the EU
Ideal use cases
- IT ticket automation
- Tool-using agents
- Long-context RAG
- Massive document analysis
- Enterprise multilingual assistants
Access & availability
Key specifications
What it costs per month
Estimate from sourced API prices and a hardware tier for self-hosting. Assumptions are editable.
Assumptions: 36-month amortisation, 0.25 €/kWh, 22 days/month, 1 $ = 0.92 €. Excludes labour, software licences, redundancy.
nAIvigate Trust Index
What the vendor offers and documents for professional use in Europe. An undocumented criterion counts as zero and stays visible: the score rises when the vendor publishes, not when we assume.
19/19 criteria documented · 100 % · 12 of them set by rule
See the criteria▾
Trust Index badge
Show this score in your docs, README or security page. It updates automatically.
[](https://naivigate.com/en/model/nemotron-3-super-120b)Data & compliance
What we know, line by line, with source and date. These lines feed the Data and Enterprise families of the Trust Index above. GDPR compliance depends on your processing, not on the model alone.
| EU data residency | Yes |
| Data processing agreement (DPA) | N/A· self-hosted: data never leaves |
| Transfer outside the EU | No |
| Training on your data | No |
| Configurable retention | N/A· self-hosted: data never leaves |
| Zero Data Retention | N/A· self-hosted: data never leaves |
| Self-hostable | Yes |
| Declared hosting | self |
Quality of this card’s data
This score measures the quality of the data on this card: price, context, Elo, licence, documented, sourced and fresh. It measures neither the model nor its compliance: for that, see the Trust Index above.
6/8 documented · 3/8 primary sources
Run it locally
| Quantization | Disk | RAM / VRAM | Typical hardware |
|---|---|---|---|
| Q4 · recommended | 68.4 GB | 77 GB | 96-128GB Mac Studio / 4× GPUs |
| Q8 · balanced | 128.4 GB | 143 GB | Server GPU infra (H100/A100…) |
| FP16 · max quality | 240 GB | 266 GB | Server GPU infra (H100/A100…) |
Estimates for a moderate context. Long contexts need more RAM (KV cache).
Deploy
Copy-ready commands generated from this card. Adjust context length and GPU count to your hardware.
OpenAI-compatible server for production on NVIDIA GPUs.
pip install vllm
vllm serve nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-BF16 \
--max-model-len 32768 \
--tensor-parallel-size 2 \
--dtype autoAdvanced data · for expertsArchitecture, modalities, detailed cost, full benchmarks▾
| Arena Elo | 1360 |
| MMLU | — |
| GPQA | — |
| HumanEval | — |
| SWE-Bench | — |
| MATH | — |