Perplexity’s local AI agent reaches Windows — with a 24GB hurdle

What this is about

Perplexity is bringing its local work agent, Portable Computer, to Windows PCs. Announced on September 14, 2026, the Windows version performs multistep tasks on the user's machine instead of sending files to a cloud model by default. Nvidia says it supports GeForce RTX and RTX PRO graphics cards with at least 24GB of video memory.

This puts an important promise of local AI against a demanding practical limit. Confidential documents can remain on the device, but the required hardware is available only in a small number of mostly expensive graphics cards. The product is therefore unsuitable for many office laptops today.

What Portable Computer actually does

Portable Computer plans work steps, searches connected files, and carries out recurring tasks. Nvidia names Qwen 3.8 27B as one example of a local model adapted for RTX hardware and Perplexity's workflows. Connectors link Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub.

The agent can organize open pull requests, check documentation for outdated sections, or analyze financial records. The companies say sensitive information stays on the machine for tasks completed locally, and those tasks do not consume Perplexity Computer credits. If local processing is insufficient, the software can use cloud models. It is supposed to ask the user for permission before sending information off the device.

Why it matters

Local processing changes the risk model. A tax document, source-code repository, or customer list does not need to be uploaded in full to an outside model provider merely for an AI system to extract information. That can simplify privacy, trade-secret protection, and internal approvals. Other data paths still exist: a user who connects Gmail, Slack, or GitHub continues to work with outside services and their permission systems.

The 24GB threshold also shows how far local agents remain from broad availability. PC Guide lists the RTX 3090, 3090 Ti, 4090, and 5090 among compatible GeForce cards. Typical corporate laptops and many gaming PCs fall below that level. This release is therefore closer to an early practical deployment for powerful workstations than a general replacement for cloud assistants.

In plain language

Think of Portable Computer as an accountant working inside your locked office. It can read and sort the files in the cabinet. Only when it needs to call an outside specialist does it ask which information may be shared. The office needs an unusually large desk, however: in this case, a GPU with at least 24GB of video memory.

A practical example

A small company has 240 monthly brokerage and tax documents covering two years. Instead of uploading every file into a web chat, it asks Portable Computer to find recurring fees locally and cite the exact file and page for every finding. If the agent identifies 18 suspicious entries, a person reviews the evidence. For a broader market comparison, the agent might then propose a cloud request. Only after approval may selected information leave the machine.

Scope and limits

• Local does not automatically mean secure: the agent receives broad access to files and connectors. Excessive permissions and malicious documents remain risks.
• Cloud escalation requires deliberate review. Users need to see which specific data will be transmitted; the launch description does not prove behavior in every edge case.
• The 24GB video-memory requirement excludes most ordinary Windows machines. Independent long-term tests of quality, energy use, and reliability are scarce at launch.

SEO & GEO keywords

Perplexity Portable Computer, Windows AI agent, local AI, Nvidia RTX, 24GB VRAM, Qwen 3.8 27B, on-device AI, privacy, cloud escalation, desktop agent