Nvidia on Thursday launched a new open-weight model designed for always-on agents and edge computing.
Nemotron 3.5 Super VL is a large 120 billion parameter model designed for enterprise agents. Although it uses fewer parameters than the Nemotron 3 Ultra model, Nvidia says it can achieve comparable accuracy in tasks that involve following instructions or using tools.
At the other end of the spectrum, the company also launched two lightweight Bolt models. These are intended to run locally on newer AI PCs (running Nvidia RTX Spark chips) and workstations running DGX Spark. These PCs are aimed at consumers, programmers, and enterprise knowledge workers, and are currently being released to the market through partners such as Microsoft and Dell.
“Organizations that want to run agents on Nvidia infrastructure now have an open, tier-optimized model that eliminates the need to choose between functionality and deployability,” said Kari Briskey, vice president of open models and software at Nvidia, in a blog post announcing these models.
Nvidia claims that when post-trained on a specific agent task, Super VL completes the task faster than competing models from Anthropic and OpenAI (Opus 5, GPT 6.1 Sol, Luna, Astra). These tasks include document intelligence, financial accounting, scientific experimentation, health insurance billing, and resiliency against security attacks.
Such claims are standard when big tech companies announce new models, but Nvidia backed it up with a list of leading companies that say they’re using Super VL. These include Skydance, EY, Ramp, Crowdstrike, Perplexity, and Salesforce.
Of course, these aren’t the first open-weight models that Nvidia has released to compete with its own Frontier Labs, a company that is both a customer and an investment in Nvidia. The GPU giant has an entire family of Nemotron models that run everything from multimodal artificial intelligence to inference, coding, math, and enterprise automation.
But these fast-growing proprietary AI labs are now looking to a number of chip and AI system manufacturers other than Nvidia (such as AWS’s Trainium and Cerebras) to fulfill their needs, and are also building their own chips.
That’s why NVIDIA is leaning aggressively into the trend of enterprise companies wanting to run their own open models in their own data centers. “Unlike the closed frontier model, organizations can train Super VL post-mortem on their own data, deploy it on their own infrastructure, and own what they build,” Briski wrote.
But with this announcement and a smaller local PC model, the company is pushing ahead with the nascent trend of consumer AI assistants. New AI assistants seem to be introduced by startups and big tech companies every day, but privacy-weary consumers may not want to invite such agents deeper into their lives and access information about their families, health, finances, and more.
So by creating machines that run models and agents locally, Nvidia hopes to create a new market for itself selling PCs with its chips, but PC users can “delegate everyday projects without sending workloads to the cloud,” Briski wrote.
Of the two Bolt models, Nemotron 3.5 Bolt 74B-A9B 24GB is a 74 billion parameter model designed for local agent tasks. It aims to balance speed and inference to handle multi-step inference, tool invocation, and web browsing. The smallest Bolt E4B is aimed at conversational agents.
Nvidia says Super VL and Bolt 74B-A9B are available now, with Bolt E4B arriving later this month.
If you make a purchase through links in our articles, we may earn a small commission. This does not affect editorial independence.
