Nvidia Nemotron 3.5: the new open models for running AI agents at work and on your PC

Nvidia Nemotron 3.5: what was announced and when

Nvidia Nemotron 3.5 is Nvidia's new family of open models built for AI agents: programs that do more than answer questions and instead carry out multistep tasks such as searching, using tools and filling in documents. The official announcement is a post by Kari Briski, Nvidia's vice president of open models and software, published on the Nvidia blog on October 15, 2026.

There was some confusion about the date. TechCrunch ran its story on Monday, October 19, 2026, saying Nvidia had launched the models "on Thursday". That Thursday is October 15, the same date shown on the Nvidia blog and on the Hugging Face model cards. So: announced on the 15th, covered by TechCrunch on the 19th.

Three models were introduced:

• Nemotron 3.5 Super VL: the large model, 120 billion parameters, for enterprise agents. Available now.
• Nemotron 3.5 Bolt 74B-A9B 24GB: a 74-billion-parameter model compressed to fit in 24GB of memory and run on your own computer. Available now.
• Nemotron 3.5 Bolt E4B 4GB: the smallest, for fast voice and conversational assistants. Coming October 23, 2026.

Nemotron 3.5 Super VL: the model for enterprise agents

Nemotron 3.5 Super VL is a 120-billion-parameter mixture-of-experts model with about 12 billion parameters active per token. In plain terms, it carries the knowledge of a large model but only runs part of it at each step, which makes it faster than a dense model of the same size.

According to the Hugging Face model card:

• it takes text, images and video as input (MP4 clips up to 2 minutes) and answers in text;
• it handles a context of up to 1 million tokens, so it can read very long documents in one go;
• it ships as a full-precision BF16 release for customization and as a compressed NVFP4 release for production use.

Nvidia describes it as a model for instruction following, tool use and multistep business workflows, with strong visual understanding: it reads text in images, tables, charts, web pages and PDFs. It is available on Hugging Face, OpenRouter, ModelScope and as an NVIDIA NIM microservice. Nvidia also released the datasets and training recipes for anyone who wants to customize it with their own data.

Nvidia and TechCrunch name Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce among the companies using it.

Nemotron 3.5 Bolt: can you run it on your PC?

Yes, Bolt 74B-A9B is built specifically to run locally, without sending your data to the cloud. It has 74 billion parameters, 9 billion of them active, and native 2-bit compression lets it fit in 24GB of memory. Nvidia designed it for the new RTX Spark laptops that arrived on October 16, but according to the blog it also runs on GeForce RTX PCs, DGX Spark, DGX Station and RTX PRO workstations. It is supported in llama.cpp, Ollama, LM Studio, Unsloth and vLLM.

The Bolt model card on Hugging Face lists a context of up to 256,000 tokens and text, image and video input, and its supported languages include Spanish, French, German, Italian and Japanese alongside English. How much memory do you really need? Here are the estimates published by Unsloth, which distributes compressed builds:

In practice, Bolt is within reach of a solid work PC or a Mac with plenty of unified memory (llama.cpp uses Metal on Macs, as the Unsloth guide explains), while Super VL needs a 64-128GB machine to run locally; otherwise the cloud is the easier route. These are estimates, not measurements: long contexts, images and video need more memory.

License and costs: what Nemotron 3.5 costs to use

Nemotron 3.5 weights are a free download and the license allows commercial use. The Hugging Face model cards list the OpenMDW 1.1 license and state that the models are ready for commercial and non-commercial use. OpenMDW is a license for AI models published by the Linux Foundation, which on May 28, 2026 described it as permissive and announced that Nvidia would adopt it for the Nemotron family. Before building it into a product, read the full license text anyway.

So what does it cost? It depends on how you use it:

• Locally: the model is free; you pay for the computer and the electricity. If you already own a machine with enough memory, the extra cost is zero.
• Through an API: Super VL is available on OpenRouter and other providers that charge per token. Rates vary by provider, so check the provider's page before you build a workflow on it.
• In a company: Nvidia distributes it as a NIM microservice, refreshed every two weeks with security patches, for teams that run it on their own servers.

Nvidia's claims: what is stated and what is not

As with every launch, many of the numbers come from the company itself. These are the figures claimed by Nvidia:

• Super VL is said to be the fastest open agentic model in its class, up to 5x faster than the next best, based on the Artificial Analysis leaderboard cited in the blog;
• it is said to match the accuracy of Nvidia's largest reasoning model with roughly one fifth of the parameters;
• Bolt 74B-A9B is said to be the top open model in its class on PinchBench and second on BrowseComp.

TechCrunch also reports Nvidia's claim that Super VL completes specific tasks (documents, accounting, medical insurance claims, security) faster than Anthropic and OpenAI models when post-trained on those tasks, but notes that no numbers or methodology were given and that such claims come with every launch. In short: interesting promises, to be checked through independent testing and, above all, on your own use cases.

It is the same race you see among closed models such as Gemini 4 Argon or Meta Muse Wonder; the difference here is that you can download the weights and keep them on your own computer.

What it means for you: creators and small businesses

For creators and small business owners, the real news is not the 120-billion-parameter model: it is that a capable AI agent is starting to fit on an ordinary computer. Here is how to make use of it:

1. Keep sensitive data in house. Contracts, invoices, client data: with a local model they never leave your machine. Nvidia's blog talks about delegating everyday projects without sending those workloads to the cloud.
2. Test before you buy. If you have a PC with at least 32GB of memory, download Bolt with LM Studio or Ollama and try it on a real task: summarize a PDF, pull data from a table, draft an email.
3. Keep your cloud tools. For images, video and avatars, open agent models do not replace creative tools; they support the organizational side of the work.
4. Read the license. OpenMDW 1.1 allows commercial use, but if you resell a service built on the model you should know its terms.

To work out which AI tools to use for creating, start with our AI guides or the guide to making an AI clone of yourself. For a complete method there is The AI Creator System, and here is how we use artificial intelligence.