NVIDIA's newest model reads text, images, documents, and video in one pass, then takes action. The company says it runs on a single desktop box.
NVIDIA ships Nemotron 3.5 Super VL for always-on agents
NVIDIA released Nemotron 3.5 Super VL on October 15, an open multimodal model aimed at AI agents that have to read a document, check facts across sources, follow a company policy, and call the right tools, all inside one request. The pitch is accuracy on messy, multi-step work rather than raw chat ability.
The model is a mixture-of-experts design, meaning it's assembled from many specialized parts and only activates the ones a given task needs. It carries 120 billion parameters in total, with 12 billion switching on per token. The context window runs to one million tokens. That combination is meant to keep the model capable without making every response crawl.
One Nemotron 3.5 Super VL checkpoint for text and visuals
Most teams building agents today stitch together separate models: one for language, another for images and documents. Context has to be handed between them, and something always gets lost in the handoff.
Nemotron 3.5 Super VL folds reasoning, instruction following, and tool use together with the ability to understand text, images, documents, and video. NVIDIA's argument is that a single model removes the plumbing. A customer-service agent can read a scanned return form and check it against policy in the same step it drafts a reply. A research agent can cross-reference a chart with a written source without a translation layer in between.
The company lists the intended jobs as coding, web research, document intelligence, cybersecurity, and video understanding. Underneath, it slots into a system of models: NVIDIA points smaller models like Nemotron 3.5 Lightning at high-volume text tasks, and keeps the larger Nemotron 3 Ultra for complex planning.
The performance claims, and who's measuring them
NVIDIA says the model delivers up to five times higher throughput on agentic tasks than other open models in its size class, when run on a single GPU. In a developer blog post, the company says it measured over 600 tokens per second on agentic tasks from the Artificial Analysis Intelligence Index, tested on a Baseten endpoint.
That's NVIDIA's own comparison, using a third party's benchmark to frame it. Independent verification of the five-times figure wasn't available at the time of writing. The speed claim rests on the compact NVFP4 weight format, a compressed version that fits the model onto a single DGX Spark, NVIDIA's small desktop AI computer. There's also a full-precision BF16 release for teams that want maximum fidelity.
Nemotron 3.5 Super VL is built to be retrained on your data
A general-purpose model won't know your return policy or your internal tool names. NVIDIA's answer is post-training: developers can fine-tune the model with NeMo, its training toolkit, using their own data, tools, and workflows.
The company offers two early examples. Scale AI, a data and technology provider, post-trained the model to automate vendor credit processing at the media company Skydance. In that workload, NVIDIA says inference cost drops 40-fold and latency falls five-fold. Financial-operations firm Ramp is using it to turn receipt images into structured data with merchant names, dates, and payment details, and expects up to 50% higher throughput.
Those figures come from the companies and NVIDIA together, so they describe what the model does in tuned, real deployments rather than out of the box.
Bolt edge models round out the Nemotron 3.5 family
Alongside Super VL, NVIDIA introduced two small models under the Nemotron 3.5 Bolt name, sized for edge devices. One is built for devices with about 4GB of memory, the other for around 24GB, which puts them in the range of laptops and small appliances.
The idea is a division of labor. Bolt handles quick, local tasks where sending data to a server is slow or unwanted. Super VL takes on the heavier multimodal reasoning. There's no single model to rule them all; NVIDIA is selling a menu.
What a 120-billion-parameter agent model means for builders
Open multimodal agent models are a crowded category now, with Meta's just-released Muse Spark Open and several strong Chinese releases competing for the same developers. NVIDIA's angle is enterprise fit: open weights, a format that runs on hardware many businesses already own, and a training path with NeMo.
For a team weighing options, the practical questions are simple. Does one multimodal checkpoint really beat a stitched pipeline in their workflows? And does the speed hold up under their own load? NVIDIA's numbers say yes. Independent testing will tell the steadier story.
The weights and the Bolt models are available now through Hugging Face and NVIDIA's developer channels. Teams already on a DGX Spark can download the NVFP4 version and start testing against their own document and video tasks.
