adtestbench Higgsfield

AI news · Nvidia · models

Nvidia releases an open model built to run AI agents all day

Nemotron 3.5 Super VL reads text, charts and web pages, and companies can retrain it on their own data and run it on their own servers.

By Alexandre S. , 00:45 UTC

Nvidia released Nemotron 3.5 Super VL on Thursday 15 October 2026, an open AI model with 120 billion parameters, the values a model learns in training, built for AI agents (programs that carry out tasks on their own) that work round the clock inside companies, Nvidia says. Any company can download the files, teach the model its own data and keep it on its own servers. VL stands for vision-language: besides text, it reads charts, tables, scanned documents and web pages.

Retro-futurist illustration: a striped sunset over a grid horizon under a starry sky, with a robot motif standing on the horizon.
Drawn by adtestbench from “NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models”,

According to Nvidia, it finishes tasks five times faster than the next best open model of its size, a claim based on Artificial Analysis results, ZDNET reports. Straight out of the box, it trails the best closed models, whose files are not public. The gains come after a company retrains it for a specific job: Nvidia cites one case of a 44 per cent improvement on medical insurance claims, ZDNET notes.

Two smaller models arrived alongside it. Bolt 74B-A9B fits in 24GB of memory to run agents on a local Nvidia machine; Bolt E4B, sized for 4GB, follows on 23 October. Salesforce, Atlassian, CrowdStrike and Perplexity are among the firms building on Super VL, Nvidia says.

Teams that want agents working on their own files, on their own hardware, now have another large US-made open model to choose from.