Nvidia has made another strategic move toward democratizing autonomous AI agents: it launched the Nemotron 3.5 family, composed of three open-weight models designed to cover everything from enterprise data centers to desktop PCs drawing just a few watts. The announcement, made on the company's official blog on October 15, marks a clear attempt to position Nvidia not just as a hardware supplier, but as a reference in software and models for the next generation of AI applications.
The most capable model in the family is Nemotron 3.5 Super VL, a 120-billion-parameter model aimed at large-scale AI agents in the corporate environment. Unlike earlier lines of Nemotron models — which already covered multimodality, coding, mathematics, and enterprise automation — the Super VL was specifically tuned for agentic tasks: following complex instructions, using external tools, executing multi-step workflows, and understanding visual documents such as PDFs, web pages, and images. According to Nvidia, it achieves 5× more throughput compared to other open models in its class on the Artificial Analysis Intelligence Index.
One of the hallmarks of the launch is that the company simultaneously released the weights, training datasets, and fine-tuning recipes. This means that companies like Skydance, EY, Ramp, CrowdStrike, Perplexity, and Salesforce can not only use the model directly but also adjust it with proprietary data, deploy it on their own infrastructure, and maintain full control over what agents do with sensitive information. In the corporate world, that capability for local post-training is often what separates a pilot project from a real deployment.
The two Bolt models, meanwhile, represent Nvidia's bet on local artificial intelligence on consumer devices. The Bolt 74B-A9B 24GB uses native 2-bit quantization to fit in just 24 GB of VRAM — enough to run on laptops with the RTX Spark chip and execute everyday agentic tasks like web browsing, tool calls, and multi-step reasoning. The Bolt E4B 4GB is even lighter, designed for local voice assistants, interactive characters, and embedded interfaces on devices with GeForce RTX or Jetson.
This market segmentation reflects a broader shift in the industry. For years, the dominant trend was to push all AI tasks to the cloud, where the scale of massive servers allowed increasingly larger models. But with the arrival of AI PCs — personal machines with dedicated AI processing chips, such as Nvidia's own new RTX Spark and DGX Spark systems — there is a growing economic and privacy incentive to run agents locally. Every token generated in the cloud costs money; every query sent to a cloud service exposes data. When an agent can complete a task in 24 GB of RAM without needing to call out to an external service, the cost per operation plummets.
Nvidia's strategy in this launch follows the pattern it has already solidified in other fronts: sell the pickaxe and the gold at the same time. The open models encourage developers to build on the Nemotron platform, and these developers will in turn need the company's hardware to run them in production. The Super VL runs on DGX Station (up to 20 petaflops of AI), the Bolt on DGX Spark or RTX Spark, and the value chain is clear. But the gain for the market is also real: Nvidia is proving that 120-billion-parameter models can compete in accuracy with 600-billion-parameter systems, provided there is better training engineering and focused post-training datasets.
What happens next is that this "desktop local agent" model could become the new standard for knowledge teams. Imagine a software engineer whose personal agent, running entirely on the laptop, organizes pull requests, reviews technical documentation, schedules meetings, and executes infrastructure commands — all without leaving the corporate network and without sending a single token outside the building. Nvidia is selling exactly that future, and partners like Microsoft, Dell, SAP, ServiceNow, and Palantir are already laying the groundwork.
The question that remains: when local agents can do 80% of what a cloud agent does, but at 20% of the cost and zero external exposure, who will still justify using cloud APIs for everyday operational tasks — and what kind of company would refuse to migrate?
Sources: TechCrunch, NVIDIA Blog, Gate News
✓ Independent sources cross-checked and verified before publishing
