Nvidia added three models to its Nemotron family on 15 October. Super VL is an open 120bn-parameter multimodal model, trained on about 26tn tokens with roughly 3.8m post-training samples and a reinforcement learning stage. Bolt 74B-A9B uses native 2-bit quantisation to fit inside 24GB of memory and is out now. A smaller Bolt E4B, aimed at 4GB, is due on 23 October.
The local-hardware angle is the interesting part. A 74bn-parameter model squeezed into 24GB runs on a single consumer card, which moves a capable agent model off rented servers and onto a desk.
The performance claims all come from one source. Nvidia says Super VL matches its own largest reasoning model on multimodal agentic tasks at roughly a fifth of the parameters, and that it is five times faster than the next best open model in its class. It puts Bolt 74B first in class on one browsing benchmark and second on another, and claims up to eight times more users at 85% lower latency when paired with its own serving software. These are vendor figures on vendor framing, not independent results. The announcement says weights, datasets and recipes ship open but does not name a licence.