NVIDIA Adds 120B Nemotron 3.5 Super VL and Two Bolt Edge Models to Open Family

NVIDIA expanded its open Nemotron lineup with three releases aimed at different points on the deployment curve: Nemotron 3.5 Super VL, a 120-billion-parameter vision-language model the company says posts the fastest task completion and highest accuracy in its class for always-on agents, plus two smaller Bolt models built for local inference. The split matters because it puts one model family on data-center racks and desktop GPUs at the same time, rather than forcing developers to stitch together separate open and closed stacks.

Super VL is the flagship of the batch. At 120 billion parameters it is sized for workloads where an agent runs continuously, handling vision and language input without a human prompting each step, and NVIDIA is claiming the top spot on task completion speed and accuracy for that specialized segment. Those two metrics pull in opposite directions in most agent deployments, since faster decoding usually costs accuracy, so the claim is effectively that Super VL's architecture narrows that tradeoff.

The Bolt models chase a different constraint: memory. Nemotron 3.5 Bolt 74B-A9B 24GB carries 74 billion total parameters but activates only about 9 billion per token, a mixture-of-experts arrangement that lets it run inside 24GB of VRAM, the ceiling on many workstation and consumer cards. NVIDIA describes its target as intelligence and speed per gigabyte for local agentic workloads, which is a distinctly different benchmark than raw accuracy on a leaderboard.

The smallest member, Nemotron 3.5 Bolt E4B 4GB, is tuned for the same local-agent role at a fraction of the memory footprint. A 4GB budget puts the model in range of laptops, mini-PCs and edge boxes where a discrete data-center accelerator is not an option. Together the three releases cover roughly three orders of magnitude in hardware requirements under a single open naming scheme.

The strategic read is straightforward. Open-weight releases have become the main way model vendors seed developer adoption, and NVIDIA's position is unusual because it sells the hardware those weights run on. Shipping a 120B flagship and 24GB and 4GB variants in one announcement means a team can prototype on a laptop and scale to a cluster without changing model family or retraining its tooling.

What to watch is whether the efficiency claims survive independent measurement. Super VL's speed and accuracy figures and the Bolt models' per-gigabyte performance are NVIDIA's own, and mixture-of-experts models in particular can behave differently under real batching and long-context loads than in vendor benchmarks. The other open question is licensing and weight availability, which determine whether these models end up inside commercial agent products or stay in research notebooks.