Nvidia has launched three open-weight AI models aimed at running agents in corporate data centers and on personal computers. The expansion of Nvidia’s open AI models pairs Nemotron 3.5 Super VL, built for enterprise tasks, with two smaller Bolt models designed to keep AI workloads on local devices rather than send them to the cloud.
Summary
Key takeaways
- Super VL has 120 billion parameters and supports enterprise agent tasks.
- Bolt targets local agents on RTX Spark PCs and DGX Spark workstations.
- Super VL and Bolt 74B-A9B are available; Bolt E4B follows later in October.
Models for enterprise agents and local PCs
Nemotron 3.5 Super VL is a 120-billion-parameter model optimized for enterprise agents. Nvidia says it delivers accuracy comparable to the larger Nemotron 3 Ultra on tasks involving instruction following and tool use, despite having fewer parameters.
The Bolt lineup serves consumers, programmers and enterprise knowledge workers. The models are designed for PCs using the Nvidia RTX Spark chip and workstations running DGX Spark. Microsoft and Dell are among the partners bringing those PCs to market.
Nemotron 3.5 Bolt 74B-A9B 24GB has 74 billion parameters. Nvidia designed it to balance speed and reasoning for local agents handling multistep tasks, calling tools and browsing the web. The smaller Bolt E4B targets conversational agents.
Together, the releases extend Nvidia’s push into always-on agents and edge computing, with models intended for both enterprise infrastructure and local AI assistants.
Performance and adoption of Nvidia’s open AI models
Nvidia says Super VL completes certain agent tasks faster than competing Anthropic and OpenAI models after task-specific post-training. Its comparisons name Opus 5, GPT 6.1 Sol, Luna and Astra.
Among the tasks addressed are document intelligence, financial accounting, scientific experimentation, medical insurance claims, and resistance to security attacks. The performance claim applies to those specialized uses, rather than a blanket comparison across all model capabilities.
Nvidia also identifies Skydance, EY, Ramp, Crowdstrike, Perplexity and Salesforce as companies using Super VL.
Enterprise customization and workloads kept off the cloud
Organizations can post-train Super VL using their own data and deploy it on infrastructure they control. Nvidia presents that flexibility as a distinction from closed frontier models.
“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,” Kari Briski, Nvidia’s vice president of Open Models and Software, wrote in the launch blog post.
Briski said organizations running agents on Nvidia infrastructure now have open, optimized models without having to choose between capability and deployability.
On PCs, the emphasis shifts to local execution. Briski wrote that users “can delegate everyday projects… without sending those workloads to the cloud.” That approach supports Nvidia’s effort to build a market for chip-powered PCs running consumer assistants locally, amid privacy concerns about agents accessing personal information.
A broader contest over AI infrastructure
Nvidia’s open AI models compete with proprietary frontier labs that are also its customers and investments. Its existing Nemotron family already spans multimodal AI, reasoning, coding, mathematics and enterprise automation.
Those proprietary labs are exploring hardware beyond Nvidia, including AWS’s Trainium and Cerebras, while also developing chips of their own. Nvidia, meanwhile, is pursuing enterprises that want to operate open models in their own data centers and consumers seeking assistants on their PCs.
According to Nvidia, Super VL and Bolt 74B-A9B are already available, while Bolt E4B is set to launch later in October.
Article produced with the assistance of artificial intelligence and reviewed by the editorial team.




