Mamba-2 + Transformer
Mamba-style layers help process sequences efficiently, while Transformer attention layers provide precise information mixing. NVIDIA combines them in a hybrid layout rather than using only one type of layer.
Nemotron 3 Super is NVIDIA’s open-weight 120B-parameter mixture-of-experts model with about 12B parameters active per token. It is built for agentic reasoning, coding, tool use and long-context tasks. This guide explains how it works, what the specifications mean, where to get it, and how to try it through Ollama or an API provider.
A plain-English introduction before looking at the technical details.
Nemotron 3 Super is an open-weight language model from NVIDIA, released as part of the Nemotron 3 family. It combines a hybrid Mamba-Transformer backbone with a Mixture-of-Experts (MoE) design. NVIDIA describes it as a model for multi-agent applications that need reasoning, coding, planning, tool calling and the ability to work with long conversations or large inputs.
The model is described as having around 120 billion total parameters, while activating around 12 billion parameters for each token. That does not make it a small model: the full checkpoint still has to be stored and served. The active-parameter figure describes the part of the model used in a forward pass, not the total storage or memory requirement.
NVIDIA released pre-trained, post-trained and quantized checkpoints, together with training datasets and recipes. Developers can use a hosted service or deploy a suitable checkpoint on their own infrastructure, subject to hardware, software and license requirements.
The values below summarize NVIDIA’s published model information. Check the exact checkpoint and serving provider before deployment.
| Specification | Published information | What it means |
|---|---|---|
| Model name | NVIDIA Nemotron 3 Super 120B-A12B | The Super model in the Nemotron 3 family. |
| Total parameters | About 120B (technical materials report approximately 120.6B) | The full model size across its experts and other components. |
| Active parameters | About 12B (technical materials report approximately 12.7B) | The approximate number used for a token’s forward pass. |
| Architecture | Hybrid Mamba-2, Transformer attention and LatentMoE | Combines sequence-processing layers with sparse expert routing. |
| Context length | Up to 1 million tokens in supported configurations | The maximum model capability is not necessarily enabled by every host. |
| Pretraining | 25 trillion tokens; NVFP4 pretraining | NVIDIA describes its low-precision training approach in the technical materials. |
| Layer and expert layout | 88 layers; 512 experts per layer, with 22 selected experts | Advanced architecture details from NVIDIA’s published training recipe. |
| MTP configuration | 2 shared-weight MTP layers | A model-side feature used for supported speculative decoding. |
| Inference feature | Multi-Token Prediction (MTP) | Can help accelerate generation through speculative decoding. |
| License | NVIDIA Nemotron Open Model License | Review the current license terms for your intended use. |
Parameter totals can be rounded differently across marketing pages, model cards and technical reports. The 120B / 12B figures are the commonly used rounded names; some technical materials specify approximately 120.6B total and 12.7B active.
The “120B” part refers to the model’s total parameter count. “A12B” means roughly 12 billion parameters are active for a given token during inference. Nemotron 3 Super uses expert routing: the model contains many expert components, but the routing system selects a subset for the current work.
This sparse design can reduce the compute used per token compared with activating every parameter on every pass. However, it does not mean the entire 120B model fits into 12B worth of memory. The weights, runtime, context cache and serving configuration still determine the hardware needed.
Its main design features are easier to understand when separated into four parts.
Mamba-style layers help process sequences efficiently, while Transformer attention layers provide precise information mixing. NVIDIA combines them in a hybrid layout rather than using only one type of layer.
Mixture-of-Experts routing sends representations to selected expert networks. NVIDIA’s LatentMoE design compresses tokens before expert processing, aiming to improve the balance between accuracy and compute cost.
MTP adds prediction heads that can propose multiple future tokens. With supported speculative-decoding implementations, this can improve generation throughput. Actual speed depends on the hardware and serving stack.
NVIDIA says Nemotron 3 Super was pretrained using NVFP4. This is a low-precision numerical format intended to improve training efficiency on supported NVIDIA hardware. It is not the same thing as a user choosing an arbitrary quantized download.
These use cases align with the model’s published focus. The right results depend on implementation, prompts, tools and evaluation.
Coordinating agents that plan tasks, hand work between tools and keep track of long interaction histories.
Supporting code analysis, debugging, planning and tool-assisted development workflows.
Working with large collections of text, lengthy technical documents and retrieval-augmented generation pipelines.
Helping automate structured triage, investigation summaries and multi-step operational tasks when connected to trusted tools and human review.
Benchmark results are useful only when the test setup is clear.
NVIDIA reports that Nemotron 3 Super achieved up to 2.2× the inference throughput of GPT-OSS-120B and up to 7.5× that of Qwen3.5-122B in a specific test using an 8K-token input and 64K-token output. NVIDIA also reports higher or comparable accuracy across a set of benchmarks and results on the RULER long-context evaluation.
Before choosing a model for production, test it on your own prompts and tasks. Measure quality, latency, throughput, tool-call reliability and total serving cost using the same conditions for each model you compare.
No. NVIDIA describes the model as supporting context lengths of up to 1 million tokens, but the context window exposed by a hosted service or local runtime can be lower. The limit is determined by that provider’s model configuration, deployment and pricing tier.
| Access route | Context shown in checked documentation | What to keep in mind |
|---|---|---|
| NVIDIA model/research information | Up to 1M tokens | Maximum model capability in a supported configuration. |
| Ollama library | 256K context shown for listed tags | Local library configuration can differ from the model maximum. |
| OpenRouter free route | 262,144 tokens shown | Free routing, rate limits and providers can change; check current availability. |
| DeepInfra endpoint | 262,144 tokens shown | Provider endpoint settings may be lower than NVIDIA’s advertised maximum. |
Provider values were checked against their public model pages on November 2, 2026. Recheck them before deployment because serving limits and free tiers can change.
Nemotron 3 Super is a large model. The hardware needed depends on the checkpoint format, quantization, context length, concurrency and inference engine. NVIDIA’s NGC model information lists 8 × H100 80GB as a minimum GPU requirement for the BF16 deployment profile. That is a reference for that configuration, not a universal requirement for every quantized or hosted version.
For a practical example, Ollama currently lists its default Super tag at about 87GB on disk and shows a 256K context window. The model file size alone is not the full runtime memory requirement: serving also needs memory for the runtime, KV cache and other processes.
Choose the route that matches your hardware and development setup.
Install Ollama first, then run the model from a terminal. The default local model download is large, so confirm you have enough storage and suitable memory before starting.
ollama run nemotron-3-superOllama may expose a smaller context configuration than the 1M model maximum. Check the current tag details before relying on long-context behaviour.
View Ollama model page ↗NVIDIA’s Hugging Face model card documents a vLLM serving path. Check the current vLLM compatibility notes and the model card’s hardware guidance before deployment.
pip install vllm
vllm serve "nvidia/NVIDIA-Nemotron-3-Super-120B-A12B-FP8"OpenRouter and DeepInfra list hosted access routes. Check the current price, context window, rate limits, data-retention terms and provider availability before sending production data.
Use NVIDIA’s research page as the main starting point for the official checkpoint list.
A quantized deployment-oriented checkpoint for supported serving stacks.
Find the available NVFP4, FP8, BF16 and base BF16 checkpoint links, technical report and supporting resources.
NVIDIA makes model weights, datasets and training recipes available for Nemotron 3 Super. It is commonly described as an open-weight model. The published model is distributed under the NVIDIA Nemotron Open Model License, so do not assume that every use is unrestricted or that the license is identical to a standard permissive software license.
Read the current license and any applicable acceptable-use terms before commercial deployment, redistribution, fine-tuning or use with sensitive information. This guide is technical information, not legal advice.
Nemotron 3 Super is NVIDIA’s open-weight 120B-total-parameter, approximately 12B-active-parameter language model. It uses a hybrid Mamba-Transformer Mixture-of-Experts architecture and is designed for agentic reasoning, coding, tool use and long-context tasks.
120B refers to the model’s approximate total parameter count. A12B means around 12B parameters are active for each token’s forward pass. It does not mean the model only needs memory for 12B parameters.
NVIDIA says the model supports up to 1M tokens in a supported configuration. Specific services may expose smaller limits; for example, the checked Ollama tag lists 256K and the checked OpenRouter and DeepInfra pages list 262,144 tokens.
The default local model is large, and the full BF16 deployment profile has substantial multi-GPU requirements. A typical laptop may not have enough memory. A hosted API or a smaller model may be more practical unless you have suitable hardware.
After installing Ollama, run ollama run nemotron-3-super in a terminal. Check the Ollama library for the current download size, context setting and supported tags before downloading.
Yes. Hosted access is listed by providers such as OpenRouter and DeepInfra. Availability, prices, context limits, rate limits and data policies vary by provider and may change.
Model weights are publicly available under NVIDIA’s Nemotron Open Model License. Whether a particular commercial use is permitted depends on the current license and applicable terms. Review those documents before deployment.
The paper is available on arXiv as “Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning.” Read the paper ↗
Use primary documentation for the latest model files, specifications, pricing, context limits and license terms.
Nemotron3AI.com is an independent informational website and is not affiliated with, endorsed by or sponsored by NVIDIA Corporation. NVIDIA and Nemotron are trademarks of NVIDIA Corporation. Provider model availability, prices and limits can change. This page summarizes public documentation and does not claim independent benchmark testing. Last reviewed: November 2, 2026. Read our Editorial Policy.
Start with NVIDIA’s official model page for the latest checkpoints, technical details and deployment resources.