Nvidia Nemotron 3.5 tiers agents: Super VL to Bolt analysis

Everyone keeps squinting at parameter counts like they are nutrition labels for intelligence. Nvidia’s latest Nemotron drop is more interesting because it admits the quiet part: agents do not all want to live in the same apartment. TechCrunch reported that Nvidia launched new open-weight models for always-on agents and edge computing, with Nemotron 3.5 Super VL aimed at enterprise use and Bolt models intended for local AI PCs and workstations. That is the architecture story hiding under the launch confetti.

The practical takeaway is not that one model wins and the others go sit sadly in the corner. It is that agent infrastructure is becoming tiered: large models for tool-heavy enterprise workflows, smaller models for local machines where latency, availability, and deployment friction matter. This is less like choosing a brain and more like fleet management, except the scooters can browse the web and occasionally summarize your quarterly planning doc with the confidence of a raccoon in a lab coat.

TechCrunch shows the split between server agents and local agents

TechCrunch reported that Nemotron 3.5 Super VL is a 120 billion parameter model built for enterprise agents, while using fewer parameters than Nemotron 3 Ultra and targeting similar accuracy in instruction-following and tool-use tasks, according to Nvidia. That matters because enterprise agents are not just chat boxes with neckties. They need to follow instructions, call tools, and survive multistep work without wandering into the semantic woods.

TechCrunch also reported that Nvidia launched two lightweight Bolt models for the other side of the deployment map. Nemotron 3.5 Bolt 74B-A9B 24GB is described as a 74 billion parameter model for local agentic tasks, including multistep reasoning, tool calling, and web browsing. The smaller Bolt E4B is aimed at conversational agents, which is a polite way of saying not every workload needs the model equivalent of a forklift.

@title Nvidia agent tiers
@source TechCrunch
@source The Next Web

  Model              Size       Target
  Super VL           120B       Enterprise agents
  Bolt 74B A9B       74B 24GB   Local tasks
  Bolt E4B           4GB        Conversation

@caption Nvidia is splitting agent models by scale and deployment target.

The key distinction is not merely big versus small. It is where the agent runs and what kind of mess it is expected to clean up. Super VL is the enterprise tier for heavier workflows, while Bolt is the deployability tier for local systems. If your procurement spreadsheet needs a model, fine. If your laptop needs a local assistant that does not phone the mothership for every sneeze, that is a different animal.

The Next Web explains why local agents are a deployment story

The Next Web reported that RTX Spark laptops are built to run AI agents on the device itself, with systems offering up to 128GB of shared memory and up to 1 petaflop of AI performance. It also reported that Nemotron 3.5 Bolt has 74 billion parameters, 9 billion active at a time, and is squeezed into 24GB of memory on supported RTX Spark configurations. That is the local-agent pitch in miniature: make the model small enough, sparse enough, or packaged well enough that it can live near the user.

This is where capability and deployability start wrestling in the parking lot. A bigger model may handle harder workflows, but a local model can be useful when the workload benefits from running close to files, apps, or users. For builders, the question becomes less romantic and more operational: where should this agent execute so it is fast enough, governed enough, and cheap enough to keep around after the demo ends?

VentureBeat and WSJ point to the enterprise reason for open models

VentureBeat reported that Nvidia says organizations can post-train Super VL on their own data, tools, and policies, then deploy it on infrastructure suited to security and data-residency needs. That is the boring sentence with the actual budget behind it. Enterprises do not just want a clever model; they want one that can be adapted to their procedures without turning compliance into a piñata.

The Wall Street Journal reported that Nvidia’s Kari Briski said interest in Nemotron has grown partly because the models are U.S.-made and because their underlying data can be inspected. WSJ also reported that public-sector organizations and financial institutions are especially cautious about Chinese-made models, and that companies are becoming more comfortable fine-tuning open models. Translation: the model choice is now also a supply-chain, governance, and customization choice. AI procurement has become less sci-fi, more paperwork goblin.

What builders should watch next, according to the launch pattern

TechCrunch framed the launch as Nvidia giving organizations an optimized open model for different infrastructure tiers, rather than forcing one choice between capability and deployability. That framing is useful if you are planning agents beyond a prototype. Put heavy, tool-using workflows on the enterprise tier; put conversational or local task agents where the hardware can actually carry them without bursting into interpretive dance.

The missing piece is independent evidence from real deployments. Nvidia’s claims are clear, and the tiering is sensible, but builders should still test accuracy, latency, tool reliability, and memory behavior on their own workloads. Watch whether teams actually route agents across cloud, workstation, and PC tiers, or whether everyone just invents a new orchestration diagram to feel alive. The agent era is not one model to rule them all, it is a dispatch problem with better GPUs and worse meeting notes.