Nvidia Pushes Open AI Models Deeper Into Corporate Servers and Home PCs

Nvidia has released a new family of open-weight artificial intelligence models built to run software agents that work continuously inside corporate data centers or directly on personal computers, extending the chipmaker's effort to stay central to the AI industry even as rivals build their own hardware and software stacks.

The centerpiece of the release, Nemotron 3.5 Super VL, is a 120-billion-parameter model intended for enterprises that want to run so-called agentic systems, programs that carry out multistep tasks largely on their own, inside their own infrastructure rather than through a subscription to a closed service. Nvidia paired it with two smaller "Bolt" models meant to run locally on new AI-equipped laptops and workstations, shifting some agent work off the cloud entirely. The company published a technical account of the model's training process and released the underlying weights through a public listing on Hugging Face, where outside researchers can inspect and modify them directly.

What Nvidia is claiming

Nvidia says Super VL matches the accuracy of its own largest reasoning model, Nemotron 3 Ultra, on multimodal agentic tasks while using roughly one-fifth the parameters, and it is marketing the model as the fastest open agentic system in its size class. The company points to the independent Artificial Analysis Intelligence Index to support a claim that it completes tasks about five times faster than the next-best open model. Super VL was trained on roughly 26 trillion tokens of data and then refined through reinforcement learning on about 3.8 million curated examples across dozens of simulated work environments. Run alongside Nvidia's Dynamo inference software, the company says the model can serve eight times as many simultaneous users with 85 percent lower latency than its predecessor.

Nvidia also told reporters that Super VL, once customized for a specific job, completes certain tasks faster than general-purpose models from Anthropic and OpenAI, naming the companies' Opus 5 and GPT 6.1 systems among the comparisons, on tasks including document review, financial accounting, scientific experimentation, insurance-claims processing and cybersecurity defense. Those comparisons come from Nvidia's own testing on narrow, post-trained versions of the model rather than from an independent benchmark, a distinction the company did not emphasize in its marketing material.

The smaller Bolt 74B-A9B model, built for local multistep reasoning, tool use and web browsing, uses native 2-bit compression to fit inside 24 gigabytes of memory, within reach of a high-end laptop rather than a server rack. Nvidia says it ranks first among open models of its size on the PinchBench benchmark and second on BrowseComp. A still smaller model, Bolt E4B, sized for voice assistants and simple conversational agents, is due on October 23. The Bolt models are designed to run on laptops carrying Nvidia's new RTX Spark chip and on DGX Spark workstations, hardware that Dell and Microsoft are beginning to ship to customers.

A push to keep the AI industry on Nvidia's terms

The release builds on the Nemotron Coalition, a group of eight outside AI labs, including Mistral AI, Perplexity, Cursor and LangChain, that Nvidia organized in March to pool data, evaluation tools and computing time on its DGX Cloud platform toward open-weight models. The arrangement gives participating labs a share in models built partly with Nvidia's infrastructure, while giving Nvidia a growing library of open systems tuned to run best on its own chips.

That strategy has taken on added urgency for Nvidia as some of its biggest customers diversify away from its hardware. Several of the frontier AI labs that once relied almost entirely on Nvidia graphics processors have begun shifting some workloads to alternatives such as Amazon's Trainium chips and systems built by Cerebras, while also designing custom silicon of their own. Open models that perform best on Nvidia's software, including its NIM microservices for deploying models inside a company's own servers, give enterprises a reason to keep building on Nvidia's stack even as they can choose from any number of underlying model providers.

The local, on-device push reflects a second pressure point: resistance from consumers and corporate security teams to handing personal or proprietary data to agents running in someone else's cloud. Nvidia is marketing the Bolt models explicitly as a way to let a PC handle everyday tasks, scheduling, document drafting, basic research, without sending that information off the machine, an argument aimed at users wary of giving AI systems access to family, health or financial information.

Who is positioned to benefit

Nvidia named a cluster of early adopters for Super VL, including the entertainment company Skydance, the professional-services firm EY, the fintech company Ramp, the cybersecurity firm CrowdStrike, the search company Perplexity and Salesforce, which plans to fold the model into its own automation products. Software makers including Atlassian, Dassault Systèmes and Synopsys are integrating the model into their platforms, while Scale AI and several smaller startups are offering to customize it for specific client industries. On the hardware side, Dell and Microsoft stand to benefit from early sales of RTX Spark-equipped machines, the first wave of consumer and workstation hardware built around the new chip.

Unlike closed frontier models, organizations can post-train Super VL on their own data, said Kari Briski, Nvidia's vice president of generative AI software, describing the appeal of a model enterprises can retrain and run on infrastructure they control rather than renting access to a proprietary system.

What happens next

Nvidia's claims about speed and accuracy have drawn coverage framing them as a direct challenge to Anthropic and OpenAI, whose business models depend on keeping their best-performing systems closed. TechCrunch, which detailed the head-to-head comparisons Nvidia presented to reporters, noted that the figures are self-reported and have not been replicated outside the company. Separate coverage of the release, including an account that examined the speed claims against the Artificial Analysis index, treated the five-times figure as Nvidia's own characterization of third-party test results rather than an independently run comparison.

The remaining model in the lineup, Bolt E4B, is due for release on October 23, and Nvidia has said it expects more hardware partners to bring RTX Spark-equipped laptops to market in the coming months as Dell and Microsoft's initial devices reach retail. The Nemotron Coalition's next joint project, a base model co-developed with Mistral AI that is expected to anchor a future Nemotron 4 family, remains in development on Nvidia's cloud infrastructure. Whether enterprises embrace locally run, open-weight agents at the scale Nvidia is betting on will depend partly on how the models perform once outside researchers, rather than Nvidia itself, have had time to test them directly.