Nemotron 3.5 Super VL is the open multimodal model NVIDIA released on October 15, 2026, in a post on its technical blog. It uses a mixture-of-experts design with 120 billion total parameters, about 12 billion active per token, and context up to 1 million tokens, handling text, images, documents and short video inside one workflow. NVIDIA positions it not as a chatbot but as an always-on enterprise agent: reading PDFs, inspecting charts, calling tools and carrying multi-step business work through under company policy. Two smaller Bolt edge models arrived the same day, stretching the Nemotron family's deployment range from data centers down to desktops and embedded devices.
Three new models, each guarding one deployment tier
| Model | Size and active parameters | Target hardware | Availability |
|---|---|---|---|
| Super VL | 120B total, about 12B active | DGX Spark, H200 and H100 class systems | October 15, 2026 |
| Bolt 74B-A9B 24GB | 74B total, about 9B active, 2-bit quantisation | RTX workstations and laptops with 24GB of memory | October 15, 2026 |
| Bolt E4B 4GB | Lightweight edition, details to follow | GeForce RTX, Jetson and other edge devices | October 23, 2026 |
NVIDIA says Super VL delivers up to roughly 5x the throughput of the next open model in its class and approaches its largest reasoning model at about one-fifth of the parameter count. Those figures come from NVIDIA's own benchmarks and a leaderboard it cites, so they should be read as vendor claims until independent testing catches up. We previously covered Nemotron taking double gold medals in maths and informatics contests; that story was about reasoning ceilings, while this one is about the efficiency of fitting a model into everyday enterprise processes.
Why documents get their own spotlight this time
The agent work that eats the most human hours in companies is rarely writing poetry; it is reading paperwork: invoices, contracts, receipts and PDFs full of tables. Super VL's pitch is to read OCR output, tables, charts and structured layouts inside a single model, removing the stitch-it-together step of converting to text first and understanding later. Early uses NVIDIA lists include Scale AI post-training a vendor credit workflow for the studio Skydance, Ramp turning receipt images directly into structured JSON, and AI21 applying it to threat-intelligence analysis. Those outcome numbers also come from NVIDIA's launch material, so they work as shortlisting clues, not as purchasing conclusions.
More than weights: the post-training kit ships too
Super VL comes as base, post-trained and NVFP4-quantised checkpoints, and companies can post-train it with NeMo on their own data, tools and policies before deploying it on their own infrastructure. NVIDIA is also releasing open datasets on Hugging Face covering tool use, finance reasoning, instruction following and vision-language work, under the OpenMDW licence, which allows commercial use, fine-tuning and redistribution. For teams that want agents to stay on their internal network, the package matters: model, training recipes and data arrive together instead of a lone weights file to figure out alone. The real test comes later, because always-on agents are judged on months of stability, cost and controllability when things go wrong, and that verdict waits for the first enterprise users to publish their own numbers.
