Nvidia releases free Nemotron 3.5 models for agents and PCs

Nvidia, the company best known for AI chips, has released new open AI models called Nemotron 3.5. One big model is for company AI agents, and two smaller "Bolt" models are made to run on your own computer. It matters because these are open-weight models, which means the model files are published so organizations can run them on their own machines instead of relying on a cloud service.

What happened

Nvidia announced the new models on its blog on October 15, 2026. The biggest is Nemotron 3.5 Super VL, a 120-billion-parameter model built for "always-on" agents. Parameters are the settings a model learns during training, and more usually means a smarter but heavier model. Nvidia says Super VL can read text, images, charts, web pages and PDFs.

TechCrunch reported that Nvidia is pushing two trends: companies running open models in their own data centers, and private agents that run on PCs. Nvidia calls the models open-weight, TechCrunch noted.

The two smaller models are the Bolt family. Bolt 74B-A9B 24GB is meant for everyday agent work on a local machine, like web research, writing and fixing code, per Nvidia. Bolt E4B 4GB is the smallest model, made for conversational agents like voice assistants and interactive characters, per Nvidia.

Key facts

• Size: Super VL has 120 billion parameters, by Nvidia's own figure, per Nvidia.
• Speed claim: Nvidia says Super VL is about 5 times faster than the next best open agent model in its class, and performs close to its largest reasoning model at about one fifth the size, per Nvidia.
• Training data: Super VL was trained on about 26 trillion tokens (small chunks of text), Nvidia said on its blog.
• Fits on a PC: Bolt 74B-A9B squeezes into 24GB of memory using a trick that stores numbers in 2 bits, per Nvidia.
• When: Super VL and Bolt 74B-A9B are out now. Bolt E4B arrives October 23, per Nvidia.
• Users: Companies Nvidia says use Super VL include EY, Ramp, CrowdStrike, Perplexity and Salesforce, per TechCrunch.

Background

Nvidia already had a Nemotron family of open models for reasoning, coding, math and business tasks, TechCrunch reported. The family also includes Lightning, released in August for heavy workloads, and Ultra, for the hardest reasoning, per Nvidia.

TechCrunch also noted that big AI labs, which buy Nvidia chips and get Nvidia investment, are turning to other chips too, such as Amazon's Trainium and Cerebras, and to their own. Nvidia says that when companies train Super VL on specific agent tasks, it finishes them faster than rival models from Anthropic and OpenAI, TechCrunch reported.

What it means for you

You can download Bolt 74B-A9B from Hugging Face and run it through popular apps like Ollama and LM Studio, per Nvidia. Nvidia says it runs on GeForce RTX PCs and is built for new RTX Spark laptops.

Our take: a model that runs on your own computer means your files stay with you and there is no per-use bill. The catch is you need a strong graphics card with plenty of memory. Speed and accuracy claims here come from Nvidia, so wait for outside tests before you rely on them.

What happens next

Bolt E4B 4GB is due on October 23, and Nvidia says the business version of Super VL gets security and speed updates every two weeks, per Nvidia.

FAQ

What is Nvidia Nemotron 3.5?

It is a new set of open-weight AI models from Nvidia. It includes Super VL, a 120-billion-parameter model for business agents, and two smaller Bolt models for PCs, per Nvidia.

Can I run Nemotron 3.5 on my own computer?

The Bolt 74B-A9B model fits in 24GB of memory and runs on GeForce RTX PCs through tools like Ollama and LM Studio, per Nvidia.

Is Nemotron 3.5 free?

Nvidia says the models have open weights and lists Hugging Face for downloads, per Nvidia. Neither Nvidia's post nor TechCrunch states the license terms, so check the license before using it for work.

Sources

• TechCrunch
• news.google.com
• blogs.nvidia.com