Nvidia has introduced its latest open models designed to run agents in enterprise settings and directly on PCs. The centerpiece is Nemotron 3.5 Super VL, a 120-billion-parameter model built for large agent tasks across enterprises. Though fewer parameters than its predecessor Nemotron 3 Ultra, Nvidia claims it matches in instruction following and tool-use accuracy. These models mark another step toward offering powerful, open-weight AI solutions.
Alongside Super VL, Nvidia launched two lighter models under the “Bolt” name intended for local deployment. These are meant for AI PCs running on the RTX Spark chip, and DGX Spark workstations—machines now entering the market via partners such as Microsoft and Dell. Bolt is split into two variants: a larger 74-billion-parameter model called Bolt 74B-A9B with 24GB of VRAM, designed for multistep reasoning, tool usage, and web browsing; and a smaller Bolt E4B focused on conversational agent applications. Super VL and Bolt 74B-A9B are already available, while Bolt E4B is expected later this month.
What This Enables
Super VL is optimized for “agentic” tasks—applications that require complex reasoning, tool invocation, and following instructions. Nvidia says in benchmarks, it outperforms models from Anthropic, OpenAI, and others (such as Opus 5, GPT 6.1 Sol, Luna, and Astra) on tasks involving document analysis, financial accounting, scientific experimentation, insurance claims, and strengthening security against attacks. Several major companies—Skydance, EY, Ramp, CrowdStrike, Perplexity, and Salesforce—are already using Super VL in such applications.
Bolt models aim to bring agent capabilities to the local machine—ideal for users wary of sending private or sensitive data to the cloud. With the appropriately powered PC, tasks can be offloaded locally, improving privacy and reducing latency. Nvidia positions this as a way to empower both consumers and enterprise knowledge workers to delegate everyday tasks to AI agents running entirely on device.
Strategic Direction and Ecosystem Implications
This release reinforces Nvidia’s push into open-weight models, letting organizations train on their own data, deploy on their infrastructure, and retain ownership of their models. Such flexibility contrasts with closed frontier models where access is more controlled.
The move also aligns with broader trends: enterprises wanting sovereignty over their AI stack, and growing demand for on-device AI agents for privacy and control. As other labs and proprietary AI players explore different chips and custom hardware, Nvidia is doubling down on its own stack from chip to model. Also, by enabling agent performance on consumer and workstation PCs, Nvidia advances a nascent category of personal AI assistants.
For businesses evaluating AI models, this could shift the calculus: you won’t need to pick between power and deployability. Super VL handles heavy agent workloads; Bolt enables local agents without recurring cloud cost or privacy exposure. It’s a strategic bet that many organizations will want to own or at least tightly control their AI agents.
Whether this pays off will depend on real-world performance, ease of deployment, and how users respond to the trade-offs of local vs. cloud agents. But if Super VL and Bolt live up to Nvidia’s claims, they could significantly reshape how enterprise and consumer AI agents are built and deployed.
