Nvidia released Nemotron 3.5 Super VL on 15 October, expanding its open-model family with a 120-billion-parameter model for work involving documents, images and tools. The company also introduced two Bolt models designed for local systems, extending the family from enterprise servers to desktop and edge hardware.
Key points
- Super VL has 120 billion parameters but activates 12 billion for each token it processes.
- Nvidia reports up to five times higher throughput on an agentic-task comparison with open models in its size category.
- Bolt 74B-A9B is designed to fit within 24GB of memory, while Bolt E4B targets a 4GB footprint.
- Super VL and Bolt 74B-A9B are available; Nvidia plans to release Bolt E4B on 23 October.
Super VL reads documents and uses tools
Super VL uses a mixture-of-experts architecture: although the model contains 120 billion parameters, it activates 12 billion for each token it processes, according to Nvidia’s developer account. Think of a filing cabinet with many drawers, only some of which need opening for a particular question. That arrangement lets the model draw on a larger collection of learned patterns without using all of them for every piece of text it generates.
The model handles text, images, documents and video alongside instructions and tool use. Nvidia says its visual training covers charts, tables, recognised text in images and multipage documents. A two-stage supervised fine-tuning process trains its language and vision components with material that includes reasoning, longer passages and examples of agents carrying out sequences of actions. For developers, the intended combination is practical: one model can read a chart, connect it to a table elsewhere in a document and produce an answer without passing the task between separate text and visual models.
Sorting receipts for expense records could bring the image and its written details into the same process, if the model extracts the merchant, date and payment information accurately. That could remove a hand-off between visual reading and text processing for that task; it would still depend on the model getting the details right.
Nvidia’s developer account reports up to five times higher throughput for agentic tasks than other open models in Super VL’s size category, describing the result as running on a single GPU. Its chart identifies a Baseten endpoint and more than 600 output tokens per second on Artificial Analysis Intelligence Index agentic tasks. The comparison concerns the measured tasks and setup, rather than the time an entire document-processing workflow would take.
VentureBeat reported that Nvidia described an Artificial Analysis measurement on four Blackwell B200 GPUs, using 100,000 input tokens and producing roughly 625 output tokens per second. The publication said the accompanying index combines 10 evaluations, including coding, science and professional work, while image inputs are evaluated separately. Speed on that index therefore sits alongside, rather than measures, the document-reading capability at the centre of this release.
Scale AI and Ramp adapt Super VL
Developers can post-train Super VL with Nvidia NeMo on their own data, tools and procedures. Nvidia says its open datasets include finance reasoning, tool use, conversational material and 45 vision-language datasets; organisations can also adapt the model’s open weights and run it in supported local, data-centre or cloud environments. The route matters for specialised agents because a model that can read a document must also learn which information a particular task calls for and how to use the available tools.
Nvidia says Scale AI post-trained Super VL for Skydance’s vendor-credit processing and EY’s merger-and-acquisition due-diligence agent. For the Skydance workload, Nvidia projects that inference cost will decrease by 40x and latency by 5x. Those figures are expectations for a particular customised workload, not measurements of the model before adaptation.
Ramp is working on extracting merchant, date and payment information from receipt images into structured data. Nvidia says Ramp expects up to 50% higher throughput and 10x lower latency after post-training. The applications differ, so each reported gain belongs to its own task and customised model.
VentureBeat described an Nvidia chart covering five specialised evaluations, with gains after post-training ranging from 9 to 44% over the base model. The publication reported that Nvidia had not supplied the chart’s task definitions, sample sizes or testing protocol. The chart compares separately customised models on separate tasks, rather than presenting a general result for the unmodified Super VL model.
Bolt models target 24GB and 4GB
Nemotron 3.5 Bolt 74B-A9B 24GB is intended for local agentic work on Nvidia RTX and DGX Spark systems. Nvidia says native 2-bit quantisation allows it to fit within 24GB of memory and run alongside other applications. Quantisation represents the model’s numerical values with fewer bits, reducing the space they occupy — much as a file can be stored in a more compact form — so a larger model can fit on hardware with a tighter memory limit.
Nvidia positions that Bolt model for activities such as web research, writing and code debugging without sending the workload to the cloud. The smaller Bolt E4B 4GB is aimed instead at conversational and physical agents on GeForce RTX, Jetson, and RTX or DGX Spark hardware, including local voice and embedded interfaces. The two memory footprints give builders different local deployment options rather than interchangeable versions of the same workload.
Super VL and Bolt 74B-A9B are available, Nvidia says. Nvidia plans to release Nemotron 3.5 Bolt E4B 4GB on 23 October.
