Nvidia’s Thursday open-weight release includes Nemotron 3.5 Super VL, a 120-billion-parameter model for enterprise agents, alongside two lightweight Bolt models. The release targets always-on agents and edge computing.
Despite having fewer parameters than Nemotron 3 Ultra, Super VL can match that model’s accuracy at following instructions and using tools, according to Nvidia.
The Bolt models are meant to run locally, either on new AI PCs with Nvidia RTX Spark chips or on DGX Spark workstations. These PCs are intended for consumers, programmers and enterprise knowledge workers, and partners including Microsoft and Dell are releasing them now.
Nemotron 3.5 Bolt 74B-A9B 24GB uses 74 billion parameters for agentic tasks handled locally. With this model, Nvidia aims to balance speed and reasoning for work involving multistep reasoning, tool calling and web browsing. Bolt E4B, the smallest model, targets conversational agents.
In the announcement post, Nvidia’s vice president of Open Models and Software, Kari Briski, explained the range: “Any organization that wants to run agents on Nvidia infrastructure now has an open, optimized model for their tier without having to choose between capability and deployability.”
Nvidia claims Super VL finishes tasks faster than Anthropic and OpenAI models Opus 5, GPT 6.1 Sol, Luna and Astra, but only when it has been post-trained on specific agentic tasks. Its examples cover document intelligence, financial accounting, scientific experimentation, medical insurance claims and security attack resilience.
Major tech companies regularly make performance claims when releasing models. Nvidia also points to a list of businesses that it says use Super VL already: Skydance, EY, Ramp, Crowdstrike, Perplexity and Salesforce.
The open-weight Nemotron models already put Nvidia in competition with proprietary frontier AI labs, covering multimodal AI, reasoning, coding, math and enterprise automation. The relationship isn’t simply one of rivals: those labs are Nvidia customers, too, and Nvidia has invested in them.
As those labs grow quickly, they’re turning increasingly to other suppliers of chips and AI systems, including AWS’s Trainium and Cerebras, while also building chips themselves. Nvidia, meanwhile, is emphasizing enterprise demand for open models that businesses can run in their own data centers.
Drawing a distinction with closed models, Briski wrote: “Unlike closed frontier models, organizations can post-train Super VL on their own data, deploy it on their own infrastructure and own what they build.”
The smaller PC models also push Nvidia into the still-young market for consumer AI assistants. Although startups and big tech companies continue to introduce assistants, users concerned about privacy may be reluctant to share information about their families, health and finances with them.
Nvidia hopes to build demand for PCs using its chips by having the machines run models and agents locally. As Briski wrote, PC users “can delegate everyday projects … without sending those workloads to the cloud.”
Super VL and Bolt 74B-A9B are available now, Nvidia says. Bolt E4B is scheduled for release later this month.
