Nvidia recently introduced new open-weight models specifically designed for always-on agents and edge computing environments.
The Nemotron 3.5 Super VL is a substantial, 120-billion-parameter model created for enterprise-level agents. While it possesses fewer parameters than the Nemotron 3 Ultra model, Nvidia states it achieves similar accuracy in tasks involving instruction following and tool usage.
On the other hand, the company also launched two lighter-weight Bolt models. These are engineered to operate locally on the latest AI PCs, which are powered by the Nvidia RTX Spark chip, and on workstations utilizing the DGX Spark. Such PCs, targeting consumers, developers, and corporate knowledge workers, are currently being released into the market through partnerships with companies like Microsoft and Dell.
“Any organization aiming to deploy agents on Nvidia infrastructure now has an open, optimized model for their specific tier without needing to choose between advanced capabilities and ease of deployment,” wrote Kari Briski, Nvidia’s vice president of Open Models and Software, in the blog post announcing these new models.
Nvidia asserts that the Super VL completes tasks more quickly than competing models from Anthropic and OpenAI (such as Opus 5, GPT 6.1 Sol, Luna, and Astra) when fine-tuned for particular agentic tasks. These tasks encompass areas like document intelligence, financial accounting, scientific experimentation, medical insurance claims processing, and enhancing security attack resilience.
While such claims are customary when major technology companies announce new models, Nvidia has supported its statements by providing a list of prominent companies that are reportedly using Super VL. This list includes Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce.
These are not Nvidia’s first open-weight models designed to compete with proprietary frontier labs—companies that are both its clients and investment targets. The GPU industry leader offers an extensive family of Nemotron models that cover a wide range of applications, from multimodal artificial intelligence and reasoning to coding, mathematics, and enterprise automation.
However, these rapidly growing proprietary AI labs are increasingly exploring a variety of other chip and AI system manufacturers beyond Nvidia to fulfill their requirements, such as AWS’s Trainium and Cerebras, alongside developing their own in-house chips.
Consequently, Nvidia is heavily investing in the trend where enterprise companies desire to run their own open models within their private data centers. Briski noted, “Unlike closed frontier models, organizations can fine-tune Super VL with their proprietary data, deploy it on their own infrastructure, and maintain ownership over what they create.”
Furthermore, with the introduction of smaller, local PC models, Nvidia is also striving to propel the nascent trend of consumer AI assistants. While new AI assistants seem to emerge daily from both startups and major tech firms, privacy-conscious consumers may be reluctant to integrate such agents so deeply into their lives, potentially granting them access to sensitive information about their families, health, finances, and more.
Thus, while Nvidia aims to forge a new market for itself by selling PCs powered by its chips—machines capable of running models and agents locally—Briski stated that PC users “can delegate everyday projects… without transmitting those workloads to the cloud.”
Regarding the two Bolt models, the Nemotron 3.5 Bolt 74B-A9B 24GB is a 74-billion-parameter model designed for local agentic tasks. Its purpose is to balance speed with reasoning capabilities for handling multistep reasoning, tool calling, and web browsing. The more compact Bolt E4B is tailored for conversational agents.
Nvidia has confirmed that Super VL and Bolt 74B-A9B are currently available, with Bolt E4B scheduled for release later this month.