NVIDIA New AI Model: Nemotron 3.5 Explained

The NVIDIA new AI model most people are asking about right now is the Nemotron 3.5 family. It is not one model. It is a group of open-weight models for chat and reasoning, image understanding, and speech-to-text, and you can download the weights instead of paying for access.

This guide explains what is in the family, who each model suits, what it costs, and how to try one safely. I have not run these models myself, so everything below comes from NVIDIA’s own pages and model cards. Where a number is NVIDIA’s own claim, I say so.

Infographic of the NVIDIA Nemotron 3.5 open AI model family: Lightning, Super VL, Bolt and Transcribe
Nemotron 3.5 at a glance: the four model groups in NVIDIA’s new open family.

What is the NVIDIA new AI model, Nemotron 3.5?

Nemotron is NVIDIA’s own line of open AI models. The 3.5 release, covered on NVIDIA’s blog and its local AI post, adds several models aimed at different jobs. NVIDIA’s goal is clear: give developers capable models they can run on their own hardware, ideally on NVIDIA GPUs.

Most of the larger models use a mixture-of-experts design. In plain words, the model has a big total number of parameters, but only a small part of it works on each word. That is why you see names like “120B-A12B”: 120 billion parameters in total, about 12 billion active at a time. Fewer active parameters usually means faster and cheaper answers.

The Nemotron 3.5 models at a glance

ModelWhat it doesSize (total / active)Best for
Nemotron 3.5 LightningText model for chat and reasoning30B / 3BTrying a model locally
Nemotron 3.5 Super VLUnderstands images and text together120B / 12BDocument and image tasks on strong hardware
Nemotron 3.5 Bolt 74B-A9BText model for local agents74B / 9B (reported to fit in 24GB of memory)Powerful local agents on a high-memory machine
Nemotron 3.5 Bolt E4BSmall text model, reported to need about 4GBAbout 4BLaptops and edge devices
Nemotron 3.5 TranscribeSpeech-to-text34B / 4.3B (per Hugging Face card)Transcription and voice apps
Nemotron 3 UltraLargest Nemotron 3 model550B / 55BData-center use

Sizes come from NVIDIA’s Nemotron developer page and the Hugging Face model cards. The 24GB and 4GB memory figures come from press coverage, so check each card before you plan hardware.

The family is wider than the table suggests. NVIDIA’s Nemotron page also lists Nemotron 3 Ultra (550B total, 55B active) and Nemotron 3 Nano Omni (30B total, 3B active), plus Nano and Super models from the earlier generation. Press coverage describes the 3.5 Bolt models as built to run AI agents locally, with the 74B-A9B version fitting in about 24GB of memory and the small E4B version in about 4GB.

Is the NVIDIA new AI model free, and where can you download it?

Yes, the weights are free. NVIDIA says its weights and training data are on Hugging Face, and models are also listed on the NVIDIA NGC catalog and on Ollama. NVIDIA describes the Nemotron license as open and commercial-friendly, but read each model’s own license before you build a product on it. If you do not want to download anything, you can test the models through NVIDIA’s hosted NIM APIs at build.nvidia.com or through OpenRouter. Hosted use may be priced or rate-limited, so check the current terms.

Key features

Open weights you can download

The models are published on Hugging Face under NVIDIA’s open license (OpenMDW-1.1). Read the license text before using a model in a commercial product, since open does not always mean “no conditions”.

Image and speech support

The Super VL model reads images as well as text, and the Transcribe model turns audio into text. NVIDIA reports a word error rate of 4.71% for Transcribe on its card. That is a self-reported figure, and one news report gave a different number, so treat it as a rough guide until independent tests appear.

Built for NVIDIA hardware

NVIDIA offers a compact NVFP4 version of Super VL, which cuts memory use on supported GPUs. NVIDIA mentions its DGX Spark desktop system, RTX PRO workstations and Jetson edge devices as targets.

Limitations to know before you start

  • Big hardware needs. The 100B-class models need server or workstation GPUs. The smaller Bolt E4B is reported to need only about 4GB, so it is the one to try on a laptop, such as the one in our HP OmniBook 5 review, if it has enough memory.
  • Benchmarks are self-reported. The scores on NVIDIA’s pages are NVIDIA’s own claims, not independent tests.
  • Hosted pricing is unclear. I could not confirm a public price for NVIDIA’s hosted API, so check NVIDIA’s site for current rates.
  • Rumours. A report mentions a “Nemotron 4”, but NVIDIA has not confirmed it, so do not plan around it.

Pricing and supported systems

The weights are free to download. Your real cost is the hardware or the cloud service you run them on. For local use, tools such as Ollama, vLLM and llama.cpp run on Linux, and Ollama also runs on Windows and macOS. Whether a given model fits your machine depends on its memory needs.

How to try the NVIDIA new AI model safely

  1. Pick one model. For a first test, Nemotron 3.5 Lightning on Ollama is the easiest starting point.
  2. Install Ollama from its official website only, not from a download link in a video or ad.
  3. Read the model page for its size and memory needs, and make sure your computer meets them.
  4. Pull and run the model from the Ollama page, then test it with a few of your own questions.
  5. Do not paste private or work data until you understand where the model runs.

Ollama is not the only route. NVIDIA also points to LM Studio and llama.cpp for running these models locally, and to a DGX Spark tutorial if you own that system.

Common problems and fixes

  • Out of memory: choose a smaller model or a more compressed (quantized) version.
  • Very slow answers: the model is probably running on your CPU. Check that your GPU is detected.
  • Download fails or stalls: large files can time out, so retry on a stable connection and check free disk space.
  • Odd or wrong answers: all language models make mistakes. Verify important facts yourself.

How Nemotron compares with other AI options

Nemotron competes with other open models, and with closed services. If you only want a chat assistant, a hosted tool is simpler; our guide to AI browsers for Windows 11 covers easy options. If you follow the fast-moving AI race, see how plans can change in our GPT-6.1 Astra story and how agents are developing in our ChatGPT agents article. The main reason to pick Nemotron is control: you can run it on your own machine and fine-tune it.

Frequently asked questions

What is the new NVIDIA AI model?

The newest release is the Nemotron 3.5 family of open-weight models for text, images and speech.

Is the NVIDIA Nemotron model free?

The weights are free to download under NVIDIA’s open license. Running them costs hardware or cloud fees.

Can I run Nemotron on my own PC?

Smaller models can run locally with tools like Ollama. The largest ones need powerful GPUs.

Is there a Nemotron 4?

One news report mentions it, but NVIDIA has not confirmed it.

Where can I download NVIDIA Nemotron models?

The weights are on Hugging Face, and NVIDIA also lists them on NGC and Ollama.

What is the smallest Nemotron 3.5 model?

Bolt E4B is the small one, reported to need about 4GB of memory.

Final verdict

Nemotron 3.5 is a solid step for people who want open models they can control, especially developers with NVIDIA hardware. For everyday users, a hosted assistant is still easier. Wait for independent benchmarks before trusting NVIDIA’s numbers, and start small with one model.

Sources: NVIDIA blog: Nemotron 3.5 Super VL, NVIDIA blog: local AI and Nemotron, Hugging Face: Nemotron 3.5 Transcribe, Hugging Face: Nemotron 3.5 Super VL, Ollama: Nemotron 3.5 Lightning.

Share this post:

Leave a Comment

Your email address will not be published. Required fields are marked *

WhatsAppFacebookXLinkedInTelegram
Scroll to Top