AI · 16 October 2026 · 5 min read
Nvidia released an open agent model it says is five times faster. The small one for machines is next week
Nemotron 3.5 Super VL is a 120-billion-parameter model, available now, for agents that read pages, charts and documents. A 4-gigabyte Bolt model, which Nvidia points at physical and conversational agents, is dated 23 October. The speed claim is Nvidia’s.
Nvidia said on 15 October that it has three new open models in the Nemotron 3.5 family. The largest of the three that it shipped immediately is Super VL, a 120-billion-parameter model for what it calls always-on agents. Nvidia says it handles instruction following, tool use, multi-step business workflows and visual understanding: a web page, a chart, a PDF. It says the accuracy is comparable with Nvidia’s own largest reasoning model at about one-fifth the parameter count. It also says Super VL is the fastest open agent model in its class, up to five times the throughput of the next. The class, and the model it beat, are Nvidia’s comparison.
With it came Bolt 74B-A9B, which Nvidia says runs ordinary agent work locally: research, writing and debugging, without sending the job to a cloud. A smaller Bolt, E4B, at 4 gigabytes, is not out yet. Nvidia dates it 23 October and says it is built for physical and conversational agents on GeForce RTX, Jetson, and its Spark machines. The jobs it names for that small model are voice assistants, characters and embedded interfaces. “Physical” is in the sentence. A robot is not.
Open weights still need a computer
Nvidia’s point, repeated from earlier open releases, is that a customer can take the weights, train them further on its own data, and run them on its own machines. That is a real difference from a model that only answers through someone else’s API. It does not remove the need for the chips the weights were built to sit on. A five-times claim on a leaderboard Nvidia cites is a reason to test the model. It is not a reason to skip the test.
The line that belongs next to a robot is the one that is still a week away. A 4-gigabyte model aimed at Jetson is the size of computer a machine might actually carry. Until 23 October it is a date. After that it is still a voice-and-interface model unless someone shows it driving a joint. Super VL, the one available today, is a document-and-tool agent. Useful, and not a hand.