Skip to content

AI & Automation1 min read

Nvidia releases open multimodal models for local hosting

Nvidia launched Nemotron 3.5 Super VL and Bolt models, allowing Swiss firms to run AI agents on private hardware.

by Fini MironFounder @ Balkaris
An abstract composition printed flat in deep charcoal, bone white, warm mid grey and bright electric lime green: four large rectangles lying flat and overlapping like sheets of paper on a table, each offset slightly to the right of the one beneath it.
Drawn for this article on our own machine. It is not a photograph.

Nvidia recently launched Nemotron 3.5 Super VL, a multimodal model featuring 120 billion parameters. The-decoder reports that this open model is designed to power enterprise AI agents that function continuously in the background. The company also introduced two Bolt models for local PCs and edge devices. These releases aim to handle diverse inputs including text, images, documents, and videos.

Why it matters

Swiss companies can now run complex agents on private infrastructure without relying on third-party APIs. This move towards local hosting addresses data residency requirements for sensitive sectors like finance or insurance. You can fine-tune the model on your own datasets to handle specific internal workflows, such as fraud screening or credit scoring, while keeping all data within your own servers.

  • Reduced operational costs by avoiding per-token fees for high-volume tasks.
  • Greater control over model outputs through custom fine-tuning.
  • Faster processing for long-running background tasks on local hardware.

What to watch

Check your existing hardware compatibility. These models are optimised for Nvidia GPUs, so your current infrastructure determines your performance gains. You should also evaluate your data privacy needs. If you require complete independence from third-party terms of service, these open weights allow for private deployment. Assess whether your specific use case requires high-throughput background tasks or local, low-latency voice assistance.

Swiss firms will deploy private agents as standard infrastructure by 2025. This requires local hardware to match the inference speed of cloud providers.

Key takeaways

  • Local hosting removes reliance on third-party APIs and per-token fees.
  • Private deployment allows for fine-tuning on internal datasets for specific workflows.
  • Performance depends on whether existing Nvidia GPU hardware meets the required inference speed.

Frequently asked

Which specific Nvidia GPU models are required to run these models at production speeds?

The source does not specify exact hardware models, only that the models are optimised for Nvidia GPUs.

Can these models handle real-time voice interaction for customer service?

The release includes Bolt models for low-latency voice assistance, but your specific hardware must support the required speed.