NVIDIA on Thursday, Oct. 15, released a family of free AI models made to power AI agents, AI that can carry out tasks on its own, such as using apps and websites, instead of only answering questions. Two are out now: Nemotron 3.5 Super VL, the large model, aimed at companies' servers, and Nemotron 3.5 Bolt, a cut-down version built to run on a personal computer. Both read images and video as well as text. A third, smaller Bolt model, meant for voice assistants, interactive characters and controls built into devices, is due Oct. 23.
Bolt rests on a way of making a big model small. Super VL has 120 billion parameters, the numbers a model tunes as it learns. NVIDIA cut that down to 74 billion for Bolt, then stored most of those numbers in a much rougher, shorter form, about 2 bits each, and trained it further to win back accuracy lost along the way. The whole model fits in 24 gigabytes of memory and can run alongside other apps. NVIDIA says its experiments show that shrinking a big model this way works better than training a small one from scratch to the same size.
Anyone can download the two models out now, and their license allows commercial use. NVIDIA is also publishing some of the data it trained Super VL on, and its training recipes. Bolt is built for the RTX Spark laptops that NVIDIA says arrive on Oct. 16, and works with popular tools for running AI models on your own computer. NVIDIA says people can hand it jobs such as researching the web, writing, and finding and fixing bugs in code without sending the work to the cloud. Super VL takes more memory: NVIDIA lists setups such as two of its H100 data center chips or one of its DGX Spark desktop AI computers.
NVIDIA says Super VL is about as accurate as NVIDIA's largest reasoning model, a model that works through a problem step by step before it answers, at roughly a fifth of the size. It says that holds for agent work such as following instructions, using software tools and making sense of images, and that Super VL is up to five times faster than the next best free model of its kind. NVIDIA says it can pull text, tables and charts out of web pages and PDFs, and names Atlassian, CrowdStrike, Perplexity and Salesforce among the companies building it into their software.
These comparisons are NVIDIA's own, made with tests and rankings it chose. In its table, Bolt scores ahead of free rivals that take up similar memory, one of them NVIDIA's own, on 10 of 13 tests. NVIDIA's speed test, run on one RTX Spark laptop, shows Bolt writing code more slowly than two of three smaller rivals, though faster than the third, and taking longer than those same two to work through a set of recorded agent tasks.
Bolt is built to fit in 24 gigabytes, but NVIDIA's speed figures come from a Surface laptop with 128 gigabytes of memory, and its announcement gives no figures for machines with less.