Nvidia launches new open models to run agents in the enterprise and on PCs 

Nvidia connected Thursday launched caller open-weight models designed for always-on agents and borderline computing.

The Nemotron 3.5 Super VL is simply a large, 120-billion-parameter exemplary geared for agents utilized by the enterprise. While it uses less parameters than the Nemotron 3 Ultra model, Nvidia says it tin present akin accuracy successful instruction-following and tool-use tasks.

At the different extremity of the spectrum, the institution besides launched 2 lightweight Bolt models. These are intended to tally locally connected the caller AI PCs (running the Nvidia RTX Spark chip) and workstations moving the DGX Spark. Such PCs, aimed astatine consumers, programmers and enterprise cognition workers, are being released close present into the marketplace done partners similar Microsoft and Dell.  

“Any enactment that wants to tally agents connected Nvidia infrastructure present has an open, optimized exemplary for their tier without having to take betwixt capableness and deployability,” Nvidia’s vice president of Open Models and Software, Kari Briski, wrote successful the blog post announcing these models.  

Nvidia claims that the Super VL completes tasks faster than competing models from Anthropic and OpenAI (Opus 5, GPT 6.1 Sol, Luna, Astra) erstwhile post-trained connected circumstantial agentic tasks. These tasks see papers intelligence, fiscal accounting, technological experimentation, aesculapian security claims, and information onslaught resilience. 

While specified claims are modular whenever a large tech institution announces a caller model, Nvidia has backed it up with a database of heavy-hitter companies that it says are utilizing Super VL. These see Skydance, EY, Ramp, Crowdstrike, Perplexity and Salesforce.  

Of course, these aren’t the archetypal open-weight models that Nvidia has released that vie with the proprietary frontier labs – companies that are some its customers and its investments. The GPU elephantine has an full household of Nemotron models that bash everything from multimodal artificial quality to reasoning, coding, math, and endeavor automation. 

However, those fast-growing proprietary AI labs are present looking to galore different chips and AI strategy makers beyond Nvidia to powerfulness their needs (like AWS’s Trainium and Cerebras arsenic good arsenic gathering their ain home-grown chips).  

So Nvidia is leaning hard into the inclination of endeavor companies wanting to tally their ain unfastened models successful their ain information centers. “Unlike closed frontier models, organizations tin post-train Super VL connected their ain data, deploy it connected their ain infrastructure and ain what they build,” Briski wrote. 

But with the announcement oregon the smaller, section PC models, it is besides trying to push forward the nascent inclination of user AI assistants. While it seems a caller AI adjunct is introduced each time by startups and large tech alike, privateness weary consumers whitethorn not privation to invite specified agents truthful profoundly into their lives giving them entree to info astir their families, health, finances and truthful on.  

So Nvidia, hopes to make a caller marketplace for itself selling PCs powered by its chips by creating machines that tally the models and agents locally, PC users “can delegate mundane projects … without sending those workloads to the cloud,” Briski wrote. 

Of the 2 Bolt models, Nemotron 3.5 Bolt 74B-A9B 24GB is simply a 74-billion parameter exemplary designed for section agentic tasks. It is intended to equilibrium velocity with reasoning to handle multistep reasoning, instrumentality calling and web browsing. The smallest Bolt E4B is aimed astatine conversational agents. 

Super VL and Bolt 74B-A9B are disposable now, with Bolt E4B coming aboriginal this month, Nvidia says.