Nvidia on Thursday released a family of open-weight Nemotron 3.5 models, led by a 120-billion-parameter Super VL model built for enterprise agents and two lightweight Bolt models designed to run locally on new AI PCs. Nvidia says Super VL matches the accuracy of its larger Nemotron 3 Ultra on instruction-following and tool use despite using fewer parameters, and that organizations can post-train it on their own data and deploy it on their own infrastructure. "Any organization that wants to run agents on Nvidia infrastructure now has an open, optimized model for their tier," wrote Kari Briski, Nvidia's vice president of open models and software.
The company claims Super VL completes agentic tasks faster than competing models from Anthropic and OpenAI, naming Opus 5, GPT 6.1 Sol, Luna and Astra, when post-trained on specific jobs such as document intelligence, financial accounting and medical insurance claims. Such benchmark claims are routine at a model launch and come from the vendor's own testing. More telling is the customer list Nvidia attached to it: Skydance, EY, Ramp, Crowdstrike, Perplexity and Salesforce are among the firms it says are already using Super VL.
The strategic logic runs underneath the product. Nvidia's largest customers are the frontier labs, which are also, increasingly, trying to reduce their dependence on it. Those labs are turning to AWS's Trainium and Cerebras, and building their own chips, to escape Nvidia's pricing and supply. By giving enterprises a capable open model they can own and run on Nvidia hardware in their own data centers, Nvidia builds demand for its silicon that does not route through the handful of labs now shopping for alternatives.
The two Bolt models aim at a different, nascent market: AI assistants that run on the device rather than in the cloud. Nemotron 3.5 Bolt 74B-A9B is a 74-billion-parameter model for local agentic tasks including tool calling and web browsing, while the smaller Bolt E4B targets conversational agents and arrives later this month. They are built for the AI PCs now reaching the market through partners such as Microsoft and Dell, running Nvidia's RTX Spark chip, which the company pitches as a way to "delegate everyday projects without sending those workloads to the cloud." The privacy argument cuts both ways: consumers wary of cloud assistants may be no more eager to give a local agent deep access to their families, health and finances.
The release extends a pattern AI Daily tracked on 18 October, when Meta and Apple both leaned harder into open-weight releases. Open weights have become the lever companies without a frontier-scale consumer product use to stay central to enterprise deployment, letting customers fine-tune and self-host rather than rent a closed model by the token. Nvidia already ships a sprawling Nemotron family spanning reasoning, coding and multimodal work; what is new is the explicit framing of enterprises owning what they build. Super VL and Bolt 74B-A9B are available now, with Bolt E4B to follow this month.