Nvidia Unveils 3 Open Nemotron 3.5 AI Models That Run on Device

Nvidia has released open Nemotron 3.5 models that run directly on devices in businesses and on personal computers. The models are designed for always-on AI agents, meaning artificial intelligence programs that perform tasks independently. The family’s flagship model, Nemotron 3.5 Super VL, has 120 billion parameters. According to Nvidia, although the model uses fewer parameters than the larger Nemotron 3 Ultra, it delivers similar accuracy when following instructions and working with various tools.

Nvidia claims that Super VL runs faster than Anthropic and OpenAI models when it is additionally trained for specific agent tasks. The tests covered document processing, financial accounting, scientific experiments and health insurance claims. The model’s resilience to security attacks was also tested. Skydance, EY, Ramp, CrowdStrike, Perplexity and Salesforce are already listed among the companies using the model.

Thanks to its open design, enterprises can further train the Super VL model on their own data and run it on their own servers. This keeps the data and the resulting system under the company’s control. Nvidia is thus targeting enterprises that want to run artificial intelligence models in private data centers. The company is expanding this strategy at a time when major AI labs are showing interest in alternatives such as AWS Trainium and Cerebras.

Two lighter Bolt models for personal computers have also been unveiled. The 74-billion-parameter Nemotron 3.5 Bolt 74B-A9B runs with 24 GB of memory. The model can complete multistep tasks, use various tools and search the internet. The smaller Bolt E4B is intended for conversational agents. Both models will run on AI PCs equipped with Nvidia RTX Spark chips and on DGX Spark workstations.

Such computers are already being brought to market with the help of partners such as Microsoft and Dell. Because computations are performed on the device itself, data related to everyday projects does not need to be sent to cloud systems. This approach is intended for users who do not want to share family, health and financial information with AI assistants. Super VL and the 74-billion-parameter Bolt model are already available. Bolt E4B will also be released by the end of this month.