Models
NVIDIA wants companies to have their own AI, and it's handing them the blueprints
Nemotron 3.5 Super VL reads documents, images and video, and companies can adapt it and run it on their own machines. NVIDIA boasts about its speed, but no one has checked it independently yet.

The essentials
- On October 15, NVIDIA released Nemotron 3.5 Super VL, an open model that companies can download, adapt with their own data and run on their servers (NVIDIA Blog, NVIDIA Technical Blog).
- Claimed by one partyNVIDIA claims it is up to 5 times faster than other open models of its size when it works as an agent, meaning it carries out tasks on its own (NVIDIA Blog, NVIDIA Technical Blog).
- Claimed by one partyAccording to NVIDIA's model card, as reported by MarkTechPost, its lowest score is 17.78 on a test of long tasks that simulates work at a bank (MarkTechPost).
- Claimed by one partyNVIDIA says Atlassian, CrowdStrike, Perplexity and Salesforce are building the model into their products (NVIDIA Blog).
Why it matters
NVIDIA no longer sells just chips. It now offers companies a model they can keep, tweak and monitor. It's the alternative to closed models that are rented over the internet. If its figures hold up, a company could have an agent that reads invoices or reports on its own machines, without sending data outside. This also suggests a way to sell more NVIDIA hardware.
If you use Perplexity, Jira or Confluence, this model may end up working behind the scenes without you noticing. And if you have an NVIDIA machine with 24 GB of memory, NVIDIA says Bolt lets you use this kind of assistant without connecting to the cloud.
The details
According to NVIDIA, Super VL is big but light to run. It has 120 billion parameters, the pieces that measure a model's size, but it only uses 12 billion at each step (NVIDIA Technical Blog). It reads text, images, documents and video, and understands tables and charts (NVIDIA Technical Blog). Besides the model's settings, NVIDIA is publishing the data and recipes it used to train it (NVIDIA Blog). MarkTechPost, drawing on the technical sheet, adds that its compressed version runs on a single DGX Spark machine (MarkTechPost).
NVIDIA has also released two small models to run on your own computer. Nemotron 3.5 Bolt 74B-A9B has been available since October 15 and fits in 24 GB of memory. Nemotron 3.5 Bolt E4B takes up 4 GB and arrives on October 23 (NVIDIA Blog). According to the company, Scale AI has already adapted Super VL for a Skydance task: managing credit to its suppliers. It estimates that the cost of that task will drop 40-fold (NVIDIA Technical Blog).
What we don't know
- Whether the speed NVIDIA advertises holds up in independent tests
- What exactly its license, OpenMDW-1.1, allows and forbids for those who use it
- Whether the cost and accuracy gains reported by customers can be repeated beyond the cases NVIDIA cites
Sources
- Open From Data Center to Desktop: NVIDIA Expands Nemotron Family With Highest Efficiency Super VL and Bolt Models
- NVIDIA Nemotron 3.5 Super VL Powers Specialized Agents with Multimodal Reasoning
- NVIDIA Releases Nemotron 3.5 Super VL, a 120B Open Multimodal Agent Model
- NVIDIA Open Models Advance Europe’s Sovereign Intelligence Across Industries
- Nvidia launches new open models to run agents in the enterprise and on PCs
Was this story useful? Yes Not really