Nvidia on Thursday announced a new family of open‑weight models aimed at helping organizations run AI agents both in the data center and on local PCs. The flagship release, Nemotron 3.5 Super VL, is a 120‑billion‑parameter model tuned for enterprise‑grade agent workloads. Alongside it, Nvidia introduced two lighter “Bolt” models, Nemotron 3.5 Bolt 74B‑A9B (74 billion parameters) and the even smaller Bolt E4B, designed to run on the company’s new AI PCs and workstations powered by the RTX Spark and DGX Spark chips.
The launch reflects Nvidia’s broader strategy to provide open models that can be customized and deployed without relying on proprietary, closed‑source frontier models. According to Kari Briski, Nvidia’s vice president of Open Models and Software, the Super VL gives organizations “an open, optimized model for their tier without having to choose between capability and deployability.” Briski also noted that, unlike closed frontier models, companies can post‑train Super VL on their own data, run it on their own infrastructure, and retain ownership of the resulting models.
What the Super VL Offers
Nemotron 3.5 Super VL positions itself as a mid‑size option within Nvidia’s existing Nemotron lineup. While it uses fewer parameters than the earlier Nemotron 3 Ultra model, Nvidia claims it delivers comparable accuracy on instruction‑following and tool‑use tasks after post‑training for specific agentic scenarios. The company says the model excels in areas such as document intelligence, financial accounting, scientific experimentation, medical insurance claims processing, and security‑attack resilience.
To back its performance claims, Nvidia cites a list of early adopters that are already using Super VL in production. These include Skydance, EY, Ramp, CrowdStrike, Perplexity, and Salesforce. The company says these organizations have found the model suitable for tasks that require a balance of reasoning ability and computational efficiency.
Bolt Models for Local AI PCs
At the opposite end of the spectrum, the Bolt models target users who want to run AI agents directly on their hardware rather than in the cloud. The Bolt 74B‑A9B model, with 74 billion parameters, is aimed at multistep reasoning, tool calling, and web browsing tasks. It is positioned as a compromise between speed and reasoning power for local agentic workloads.
The Bolt E4B model is even smaller and is intended for conversational agents that can handle everyday queries without needing to send data to external servers. Nvidia says the Bolt E4B will be available later this month, while the Super VL and Bolt 74B‑A9B are already accessible to developers and enterprise customers.
Briski highlighted the privacy benefits of running models locally: users can “delegate everyday projects … without sending those workloads to the cloud.” This approach appeals to consumers and enterprise knowledge workers who are wary of exposing personal or corporate data to third‑party services.
Why Nvidia Is Pushing Open Models
Nvidia’s move comes as several leading AI labs, including Anthropic and OpenAI, begin to explore alternatives to Nvidia’s GPUs, such as AWS’s Trainium, Cerebras systems, and in‑house chips. By offering open‑weight models that can be fine‑tuned and deployed on Nvidia hardware, the company seeks to lock in customers who want control over their AI stacks.
The Nemotron family already includes models for multimodal understanding, coding, math, and general enterprise automation. The addition of Super VL and the Bolt line expands the portfolio to cover both high‑parameter enterprise agents and lightweight, on‑device assistants. This breadth lets Nvidia address a wider range of use cases, from large‑scale data‑center deployments to personal AI PCs rolling out through partners like Microsoft and Dell.
Availability and Ecosystem
Developers can obtain the Super VL and Bolt 74B‑A9B models immediately through Nvidia’s usual channels, including the NGC catalog and the Nemotron research site. The Bolt E4B model will follow later in October. Nvidia also provides documentation and example scripts for post‑training the models on custom data, enabling organizations to tailor the models to specific workflows such as claims processing or internal knowledge‑base queries.
The company’s blog post announcing the release includes links to the model cards and benchmark results that compare Super VL against competing offerings from Anthropic and OpenAI (referencing Opus 5, GPT 6.1 Sol, Luna, and Astra). While such performance claims are typical for a major hardware vendor’s model launch, Nvidia supports them with named enterprise customers who are already testing or deploying the models in real‑world settings.
Looking Ahead
The release signals Nvidia’s confidence that enterprises will continue to favor open, controllable AI solutions, especially as data privacy regulations tighten and organizations seek to avoid vendor lock‑in. At the same time, the push toward local AI agents on PCs could help grow a new market segment for AI‑enabled hardware, encouraging consumers to adopt assistants that keep personal information on‑device.
For IT leaders and developers, the immediate next steps involve evaluating whether the Super VL’s parameter size and licensing terms fit their existing infrastructure, experimenting with post‑training on domain‑specific data, and assessing the Bolt models for edge or end‑user scenarios. As more companies share their experiences with these models, the broader impact on enterprise AI adoption and on‑device AI usage will become clearer.
Filed under: AI, TechNews, Startups, ProductLaunches, Software