NVIDIA released Nemotron 3.5 Super VL on October 15, 2026, an open 120-billion-parameter model aimed at specialised always-on agents. The company says Super VL matches the accuracy of its largest reasoning model on multimodal agent tasks while using about one-fifth the parameters, and ranks it as the quickest open agent model in that band on the Artificial Analysis Intelligence Index, citing up to 5x the throughput of the next peer on that board.
Super VL was built with the Nemotron Coalition and targets instruction following, tool use, multi-step business workflows and visual understanding. Because the weights are open, NVIDIA emphasises that organisations can post-train the model on their own data, run it on their own infrastructure and keep oversight of what they deploy, in contrast to closed frontier APIs.
The same post expands the edge side of the Nemotron line. Nemotron 3.5 Bolt 74B-A9B 24GB is available now for local agent work on NVIDIA RTX and DGX Spark systems using native 2-bit quantisation to fit in 24 GB. A smaller Bolt E4B 4GB model for conversational and embedded agents is scheduled for October 23. An earlier Bolt 74B announcement from October 12 already appeared in published news; this package centres on Super VL as the new server-class release while treating Bolt details as accompanying context.
NVIDIA places Super VL beside Lightning for high-throughput work and Ultra for the hardest reasoning jobs, with Bolt covering the desktop and edge. Together the company presents a tiered open stack for agent deployments from data centre to device. Super VL and Bolt 74B-A9B are available immediately according to the blog.
Adoption colour in the post includes organisations already post-training Super VL. NVIDIA says Scale AI post-trained the model for Skydance's vendor credit finance agent and for EY's M&A due diligence agent, and that AfterQuery customized it for Hebbia credit-agreement analysis. Those examples are vendor-supplied customer stories, not independent case studies audited here.
Performance claims such as 5x throughput and PinchBench or BrowseComp rankings for Bolt variants come from NVIDIA's announcement and cited boards. This package records them as company-reported figures. What is firmly established is the product event: an open 120B multimodal agent model is shipping now under the Nemotron 3.5 Super VL name.
NVIDIA also ties Super VL to a broader message that open agent models should span the full deployment ladder, not only the largest training clusters. For teams already committed to NVIDIA agent tooling, the release reduces the need to choose between a closed frontier API and a smaller open checkpoint that cannot hold multimodal workflows.
For builders, the practical shift is having an open, post-trainable multimodal workhorse sized for always-on agents without jumping to NVIDIA's largest closed reasoning tier. Remaining unknowns include full license text details beyond the open-model framing, exhaustive third-party replications of the Artificial Analysis comparison, and how Super VL pricing and support land across cloud partners.