Running AI agents locally: Perplexity and NVIDIA bet on on-device data processing

21 September 202616 views

Companies introduced Portable Computer — an environment where agents, tools, and code execution run on the user's own hardware, with data transferred to the cloud only after their explicit consent. Currently, the solution is designed for Linux with NVIDIA GPUs of at least 24 GB VRAM, including DGX Spark, while Windows support is promised for a later date.

Running AI agents locally: Perplexity and NVIDIA bet on on-device data processing

What Perplexity and NVIDIA Announced

Perplexity and NVIDIA introduced Portable Computer. It is a local AI tool for running agents on the device.

  • Portable Computer runs AI agents directly on devices.
  • The tool gives greater control over data.
  • It requires powerful NVIDIA hardware and Linux support.
  • Portability allows working without relying on cloud services.
  • A Windows version is expected.

The partnership reflects the trend toward local AI processing. Conclusion: Portable Computer is suitable for local work, but depends on hardware.

How Local Processing Protects Data

Local execution reduces the need to send data to the cloud. Information stays on the device if the task does not require cloud services.

  • The device executes code and uses various tools.
  • The application requests permission to upload to the cloud.
  • The user chooses whether to upload data.
  • The approach is relevant for personal and professional data.

Conclusion: local processing gives control over data, but does not eliminate the cloud.

Hardware and Platform Requirements

Portable Computer requires Linux. An NVIDIA GPU with at least 24GB VRAM is needed.

  • This is the main limitation of the tool.
  • The tool is suitable for professionals and users of high-end systems.
  • Currently, Portable Computer runs on NVIDIA DGX Spark.
  • Soon it will run on NVIDIA RTX GPU PCs.
  • Windows support is expected.

Conclusion: advanced NVIDIA hardware is required to run it. The tool targets users with hardware for demanding models.

Models and Access Conditions

The system supports several models. The first release is available to Pro and Max subscribers on NVIDIA DGX Spark.

  • Qwen 3.8 27B
  • PPLX 27B
  • NVIDIA Nemotron 3.5 Lightning will appear in the model selection.

Conclusion: a selection of models already exists, but access is limited by subscriptions and hardware.

Local Execution or Cloud: Selection Criteria

Comparing local and cloud approaches helps in choosing a strategy. Portable Computer runs on the device.

CriterionPortable ComputerTraditional cloud AI services
Where tasks are executedOn the user's deviceOn cloud servers
Control over dataThe user chooses whether to upload to the cloudCloud services require uploading data
Hardware requirementsLinux, NVIDIA GPU, at least 24GB VRAMSuitable for users without powerful NVIDIA hardware
AvailabilityPro and Max subscribers on NVIDIA DGX Spark; Linux firstBroader for users without powerful hardware

Conclusion: Portable Computer is suitable for confidential files and powerful hardware. For others, cloud AI services are more practical.

Frequently asked questions

Running AI agents locally: Perplexity and NVIDIA bet on on-device data processing