Skip to content
NewsModel release

Nvidia ships Nemotron 3.5 Super VL and Bolt models for on-device AI agents

· by Pondero Newsdesk

The short version

Nvidia released a 120-billion-parameter open agent model plus a 74-billion-parameter sibling that squeezes into 24GB of memory, aiming to run enterprise and desktop agents without a cloud round trip.

Nvidia ships Nemotron 3.5 Super VL and Bolt models for on-device AI agents

Nvidia's newest open model fits a 74-billion-parameter agent inside 24GB of memory using 2-bit quantization, a compression step aimed at letting multistep agent work run on a desktop instead of a cloud API.

What

Two models landed on October 15, 2026: Nemotron 3.5 Super VL, a 120-billion-parameter (12 billion active) open vision-language model built for always-on enterprise agents, and Nemotron 3.5 Bolt 74B-A9B, a 74-billion-parameter local model squeezed into 24GB of memory per Nvidia's announcement. A smaller sibling, Bolt E4B at 4 billion parameters and aimed at conversational and device agents, arrives October 23 per Nvidia. On agentic tasks, Nvidia claims Super VL matches its much larger Nemotron 3 Ultra reasoning model at roughly one-fifth the parameter count, running up to 5x faster than the next-best open agentic model in its class per Nvidia's blog post. Outside testing backs part of that up without matching the superlatives: Artificial Analysis scored Super VL at 26.5 and Bolt at 21.4 on its Intelligence Index per Artificial Analysis, as reported by Let's Data Science. The timing lines up with hardware. Bolt 74B-A9B targets Nvidia's new RTX Spark and DGX Spark machines, which began reaching PC makers including Microsoft and Dell earlier this month per TechCrunch.

A local-only deployment tier opens for agent builders

Bolt 74B-A9B gives teams that cannot route sensitive tasks through a third-party API a way to run multistep tool use, coding, and web-research agents entirely on local hardware instead. Nvidia's own examples lean toward regulated work: insurance claims review and credit-score checks where the model, the data, and the decision never leave the premises per Nvidia's blog post. A longer adopter list backs the enterprise side of the release. Atlassian, CrowdStrike, Perplexity, and Salesforce are among the companies Nvidia says are post-training or integrating Super VL today per Nvidia, alongside Scale AI's work building agents for Skydance and EY. None of that is independently confirmed; it is Nvidia's own list of named customers, not a third-party audit. What matters for buyers is whether those integrations reach production, not a pilot slide.

What to watch next

Bolt E4B's October 23 release offers the first outside look at Nvidia's smallest agent model per Nvidia. After that, independent tool-use and multistep-browsing benchmarks should start surfacing as developers get broader hands-on access to Bolt 74B-A9B and Super VL.

Sources