PickNVIDIA ships Nemotron 3.5 Super VL and Bolt
NVIDIA released Nemotron 3.5 Super VL, an open 120B multimodal model for always-on agents, plus Bolt edge models for local RTX and DGX Spark systems. Super VL and Bolt 74B-A9B are available now on Hugging Face; Bolt E4B follows Oct. 23.
NVIDIA has added three open models to its Nemotron family, each sized for a different place an AI agent might run: the data center, a local workstation, or a small edge device.
The lineup
Nemotron 3.5 Super VL is an open, 120-billion-parameter multimodal model built for what NVIDIA calls specialized, always-on agents. The company says it matches its own largest reasoning model on multimodal agent tasks such as instruction following, tool use, multi-step business workflows and visual understanding, while using roughly a fifth of the parameters. NVIDIA also claims it is the fastest open agentic model in its class.
Alongside it come two edge models. Nemotron 3.5 Bolt 74B-A9B uses native 2-bit quantization to fit in about 24 GB of memory, so it can run everyday agent tasks locally on RTX and DGX Spark systems next to other applications. Nemotron 3.5 Bolt E4B is a much smaller 4 GB model aimed at fast conversational agents.
Super VL and Bolt 74B-A9B are available now on Hugging Face. Bolt E4B is scheduled for October 23.
Why it matters for builders
The interesting part is not any single model but the tiering. Many teams want the same agent logic to run in the cloud for heavy jobs and locally for private or latency-sensitive ones. A family with shared behavior across sizes makes that easier: you can prototype against the big model, then test whether a smaller one handles the routine 80 percent of tasks.
Open weights also mean you can post-train on your own data and deploy on your own infrastructure, which matters for companies that cannot send documents or screenshots to a third-party API. NVIDIA explicitly pitches this as an advantage over closed frontier models.
The caveat is that "fastest in class" and accuracy comparisons are vendor claims. Before switching, run the model on a sample of your real agent traces, including tool calls and image inputs, and compare cost per completed task rather than raw benchmark scores. Local options like Ollama and hosted routers like OpenRouter make that kind of side-by-side test straightforward.
Nemotron 3.5 Super VL is listed in our New AI Models category.
Source: NVIDIA. This article is PickAI’s own summary and analysis based on the original reporting.
