Nvidia unveiled a set of new open-weight models on Thursday, built for always-on agents and edge computing scenarios. The Nemotron 3.5 Super VL is a 120-billion-parameter model aimed at enterprise-grade agents. Although it carries fewer parameters than the Nemotron 3 Ultra, Nvidia maintains it achieves comparable accuracy on instruction-following and tool-use tasks. On the lighter end, the company introduced two compact Bolt models meant to run locally on the new AI PCs powered by the Nvidia RTX Spark chip, as well as workstations built on DGX Spark. These consumer-oriented machines—targeting everyday users, programmers, and enterprise knowledge workers—are hitting the market right now through partners including Microsoft and Dell.

"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," wrote Kari Briski, Nvidia's vice president of Open Models and Software, in the blog post announcing the releases. Nvidia asserts that Super VL finishes tasks faster than rival models from Anthropic and OpenAI (Opus 5, GPT 6.1 Sol, Luna, Astra) when post-trained for specific agentic workloads. Those workloads span document intelligence, financial accounting, scientific experimentation, medical insurance claims, and security attack resilience. Such performance claims are routine whenever a major tech player launches a new model, but Nvidia reinforced its assertion with a roster of prominent companies it says are using Super VL—among them Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce.

These are hardly Nvidia's first open-weight models competing with the proprietary frontier labs—companies that are simultaneously its customers and its investments. The GPU giant already fields an entire Nemotron family covering everything from multimodal AI to reasoning, coding, math, and enterprise automation. Yet those fast-growing proprietary AI labs are increasingly turning to other chip and AI system providers beyond Nvidia—such as AWS's Trainium, Cerebras, and their own in-house silicon—to meet their compute needs. In response, Nvidia is betting heavily on the enterprise appetite for running open models inside companies' own data centers. "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," Briski wrote.

With the smaller, local PC models, Nvidia is also trying to accelerate the still-nascent consumer AI assistant trend. New AI assistants seem to launch daily from startups and big tech alike, yet privacy-conscious consumers may hesitate to let such agents deeply into their lives—granting access to details about their families, health, finances, and more. Nvidia hopes to carve out a new market for itself by selling PCs powered by its chips that run models and agents locally, so that PC users "can delegate everyday projects … without sending those workloads to the cloud," per Briski.

Among the two Bolt models, the Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model built for local agentic tasks, balancing speed with reasoning to handle multistep reasoning, tool calling, and web browsing. The smallest, Bolt E4B, targets conversational agents. Super VL and Bolt 74B-A9B are available now, while Bolt E4B arrives later this month, according to Nvidia.