Every Nemotron AI Model Explained and Compared (Nov, 2026)

Nemotron Nano vs Super vs Ultra

The three size names have held since NVIDIA announced Nemotron 3 on December 15, 2025. Nano is the small, fast tier for single tasks and sub-agents. Super is the mid tier for multi-agent work on a few data center GPUs.

Ultra is the largest, for the hardest reasoning.

All of them are Mixture-of-Experts models, so only a slice of the weights runs on each token: the active count, not the total, sets the speed. The 3.5 generation then added Lightning, Super VL and Bolt, and left Ultra on version 3.

Nemotron 3 Ultra

NVIDIA pretrained Ultra in its own 4-bit NVFP4 format. The model card lists a minimum of 8x B200, 8x H200 or 16x H100 for the BF16 weights, and the NVFP4 version halves that to 4x B200 or 8x H100.

It adds Multi-Token Prediction layers that draft several tokens at once.

Its post-training data runs to May 2026, while its pretraining data stops at September 2025.

Ultra's standout number is on hallucination: NVIDIA reports 78.7 on the AA-OmniScience non-hallucination score, where the Chinese open models it compares against score between 2.8 and 74.4.

Nemotron 3.5 Super VL

Super VL is the newest large Nemotron and the one I would test first. It adds a C-RADIO vision encoder to the Super backbone, so it reads screenshots, documents and video frames and writes text.

NVIDIA's launch post says NVIDIA trained it on about 26 trillion tokens and calls it comparable to Ultra at about one-fifth of the parameters.

The model card backs part of that: Super VL scores 89.5 on GPQA-Diamond against Ultra's 87.0, but trails it on IFBench, 70.9 to 82.9. It deploys on 4x H100, half the GPUs the March Super needed.

Nemotron 3 Super

The March 11, 2026 Super is text only, with the same 120B total and 12B active split and a 1M window. Its card names IT ticket automation, tool use and RAG as target jobs and sets a minimum of 8x H100-80GB.

It is the cheapest large Nemotron to rent, at $0.075 input on OpenRouter, and it stays on the older Nemotron Open Model License.

For new text-only work, Super VL does the same jobs and adds vision; keep Super where you already tuned prompts against it.

Nemotron 3.5 Lightning

Lightning, released August 11, 2026, replaced Nano 30B at the small end. It has 30B total and 3B active parameters, runs on one H100 or A100 80GB with a 256K window, and reaches 1M on bigger setups.

NVIDIA claims up to 4x faster output than comparable models.

It is the model NVIDIA pairs with NeMo Switchyard, an open source router that sends each step of an agent to the cheapest model that can handle it.

Lightning is also the most downloaded 3.5 model on Hugging Face, past 1 million downloads in the last month for the BF16 weights alone.

Nemotron 3 Nano: 30B, 4B and Omni

Nano 30B was the first Nemotron 3 model, out December 15, 2025. Its card describes 23 Mamba-2 and MoE layers plus 6 attention layers, 128 experts with 6 active per token, and 3.5B active parameters.

NVIDIA said it delivers up to 4x the throughput of Nemotron Nano 2. Two more Nanos followed. Nano 4B (March 16, 2026) is a dense model compressed from Nano 2 9B, with a 262K window.

Nano Omni (April 28, 2026) takes video, audio, images and text, and its NVFP4 build runs on a single RTX 5090 with 32 GB.

Nemotron 3.5 Bolt 74B and E4B

Bolt is the local tier. Bolt 74B-A9B is Super VL cut down with NVIDIA's Star Elastic compression, from 22 active experts per token to 9, then stored in 2-bit weights to fit a 24 GB card.

It keeps vision input and a 256K window, and ships with llama.cpp, Ollama and LM Studio support.

Bolt E4B (October 23, 2026) is the odd one out: NVIDIA post-trained it from Google's Gemma 4 E4B, so it is a Gemma model under the Nemotron name. It fits in 4 GB and targets voice and device assistants.

What the Nemotron models can do

Every current chat model in the family follows the same pattern: it writes a reasoning trace, then a final answer, and it calls tools for agent work. The architecture is the part that differs from Llama and DeepSeek.

Most layers are Mamba-2, a state space design whose memory use does not grow with context the way attention does, so a 1M token prompt costs less to hold in memory. A few attention layers keep recall accurate.

On Ultra, NVIDIA reports 94.7 on the RULER test at 1M tokens.

Openness goes further than weights. NVIDIA released three trillion tokens of pretraining, post-training and reinforcement learning data with Nemotron 3, and each card links the datasets used.

That matters if you need to show an auditor what data a model learned from, or if you want to repeat NVIDIA's post-training on your own domain with its NeMo tools.

Context windows and long inputs

Ultra, Super VL, Super, Lightning and Nano 30B all accept up to 1M tokens. Bolt 74B and Nano Omni stop at 256K, Nano 4B at 262K, and Bolt E4B at 64K in and 64K out.

The card limit is not what you get from a host: OpenRouter's paid listings for Ultra, Super, Lightning and Nano 30B cap context at 262,144 tokens, while the free endpoints advertise 1M. Lightning's own card uses 256K for a single H100 deployment.

If you need the full window, plan to self-host on enough GPU memory, or pick a provider that lists 1M, as several do for Super VL.

Nemotron vs Llama and DeepSeek

Against DeepSeek, Ultra trades blows on NVIDIA's own scoreboard.

The Ultra card puts it ahead of DeepSeek V4-Pro on Terminal Bench 2.1 (56.4 to 49.2) and far ahead on not hallucinating, behind on SWE-Bench Verified (70.7 to 74.5), and level on long context.

These are NVIDIA's runs, not an independent test. DeepSeek V4-Pro is the bigger model at 1.6T total and 49B active, and its official API charges $1.32 input and $3.96 output at peak.

Our DeepSeek model guide covers its off-peak rates and MIT license.

Against Llama, the history matters.

Nemotron started as Llama fine-tunes: Llama-3.1-Nemotron-70B in October 2024, then Llama-3.3-Nemotron-Super-49B and Llama-3.1-Nemotron-Ultra-253B in 2025, each built by compressing and retraining a Meta model.

Nemotron 3 broke from that with NVIDIA's own pretrained hybrid architecture. Meta, for its part, has shipped no new Llama since April 2025 and now releases open weights under the Muse name, as our Llama model guide explains.

If you run Llama 4 today, Super VL is the closest Nemotron swap: similar active size to Maverick, vision input, and a license with no 700 million user threshold.

My read: pick Nemotron when you want open training data, a permissive license with no user cap, or a model tuned for NVIDIA GPUs and NIM. Pick DeepSeek when raw benchmark scores per dollar on its own API matter most.

Llama 4 is still cheap to rent, but it has stopped moving.

Older Nemotron models

NVIDIA leaves older Nemotron weights on Hugging Face rather than deleting them, so old checkpoints keep working if you self-host. The timeline shows how the family moved from Llama derivatives to its own architecture.

Moving from a Llama-based Nemotron to Nemotron 3 is more than a model swap.

The architecture, the chat template and the thinking switch all differ, and the hybrid layers need a recent vLLM or SGLang build with Mamba support, so budget a re-test of prompts and serving config.

The Super 49B v1.5 card documents the old /no_think behaviour if you need to compare.