Nvidia has introduced a new set of open-weight AI models built to support always-on intelligent agents, targeting both large-scale enterprise deployments and consumer-grade personal computers running locally on-device hardware. The announcement, made on Thursday, spans a broad range of computing environments - from data centers to AI-enabled laptops and workstations.

The flagship release is the Nemotron 3.5 Super VL, a 120-billion-parameter model developed with enterprise-grade agentic workloads in mind. Despite having a smaller parameter count than its predecessor, the Nemotron Ultra, Nvidia states that Super VL achieves comparable accuracy on instruction-following and tool-use benchmarks. The company is positioning it as a capable, deployable option for organizations that need powerful AI agents without the overhead of the largest available models.

Lightweight Bolt Models Target Local AI on Consumer Hardware

On the opposite end of the scale, Nvidia simultaneously unveiled two compact models under the Bolt family name. These are engineered to run directly on local hardware - specifically the new generation of AI PCs equipped with Nvidia's RTX Spark chip, as well as workstations powered by the DGX Spark platform. These machines are currently entering the consumer market through hardware partners including Microsoft and Dell, targeting everyday users, software developers, and enterprise knowledge workers who prefer on-device processing.

Kari Briski, Nvidia's Vice President of Open Models and Software, framed the breadth of the release in a company blog post. "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," she wrote, signaling that the lineup is intended to cover every level of the computing stack.

Performance Claims and Enterprise Adoption

Nvidia is making strong performance assertions for Super VL, claiming it completes agentic tasks faster than comparable offerings from Anthropic and OpenAI - specifically citing models such as Opus 5, GPT 6.1 Sol, Luna, and Astra - when the model has been post-trained on domain-specific tasks. The use cases highlighted by Nvidia include document intelligence, financial accounting, scientific research workflows, medical insurance claim processing, and cybersecurity resilience against attacks.

Such performance claims are a routine feature of major AI model announcements. However, Nvidia has sought to add credibility to its assertions by disclosing a list of prominent organizations it says are already deploying Super VL. That roster includes Skydance, EY, Ramp, CrowdStrike, Perplexity, and Salesforce - names that carry significant weight across entertainment, financial services, cybersecurity, and enterprise software sectors.

A Broader Open-Model Strategy

These releases are not Nvidia's first foray into the open-weight model space. The company has been steadily building out its Nemotron model family, which already covers a wide range of AI capabilities including multimodal processing, reasoning, coding, mathematics, and enterprise automation. What is notable about this latest batch is the strategic context in which it arrives.

The major proprietary AI laboratories - companies that are both customers and investment targets for Nvidia - are increasingly diversifying the chips and infrastructure they rely on. Alternatives such as AWS's Trainium processors, Cerebras systems, and in-house custom silicon developed by the labs themselves are gaining ground. This shift creates pressure on Nvidia to establish stronger footholds in adjacent markets.

One of those footholds is the growing appetite among enterprises to operate their own AI models within their own infrastructure, rather than relying on closed, externally hosted systems. Briski addressed this directly: "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." This positions Super VL as a tool that gives companies genuine ownership over their AI capabilities - a consideration that is becoming increasingly important for regulated industries and data-sensitive organizations.

The Push Toward On-Device Consumer AI

The Bolt model announcements reflect a different but related ambition - advancing the still-emerging category of consumer AI assistants that operate entirely on local hardware. The AI assistant market has grown increasingly crowded, with new products arriving from both startups and established technology companies on a near-daily basis. Despite this activity, a meaningful portion of consumers remain wary of granting AI agents deep access to personal information spanning family life, health data, and financial records.

Nvidia's response to this concern is to promote local processing as a privacy-preserving alternative to cloud-based AI. By running models directly on the device, users avoid transmitting sensitive data to remote servers. As Briski described it, PC users "can delegate everyday projects without sending those workloads to the cloud." In doing so, Nvidia is also working to create a new commercial market for itself - one built around selling AI-capable PCs powered by its own chip architecture.

Model Specifications and Availability

Among the two Bolt models, the Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model built for local agentic tasks. It is designed to balance processing speed with reasoning depth, making it suitable for multistep reasoning tasks, tool calling, and web browsing operations. The second model, the Bolt E4B, is the smallest in the lineup and is targeted at conversational agent applications where responsiveness and efficiency take priority over complex reasoning chains.

In terms of availability, Nvidia confirmed that Super VL and the Bolt 74B-A9B are accessible immediately. The smaller Bolt E4B is expected to become available later in the current month. Together, the three models represent Nvidia's most comprehensive push yet into open-weight AI for both enterprise infrastructure and consumer computing environments.