Agents

Perplexity brings Portable Computer to Windows RTX PCs

Perplexity has launched its local-first AI agent, Portable Computer, on Windows PCs with NVIDIA RTX GPUs, allowing developers to run complex workflows locally without per-token API fees.

AlphaSignal2 days agoAgents
Image: AlphaSignal

Perplexity has expanded its local-first AI agent, Portable Computer, to Windows PCs equipped with NVIDIA RTX GPUs. Previously available on NVIDIA DGX Spark and Linux systems, the tool is now accessible to Perplexity Pro subscribers paying $20 per month and Max subscribers paying $200 per month through the official Windows application. The system runs its core orchestrator, planner, models, and local search index directly on the user's hardware.

To run the agent locally, users need an NVIDIA RTX GPU with at least 24 GB of VRAM, such as an RTX 5090 or a compatible RTX PRO card. Portable Computer ships with optimized models, including Qwen3.5 27B and Perplexity's post-trained PPLX 27B, with NVIDIA's 30B Nemotron 3.5 Lightning model expected to arrive soon. While Qwen3.5 27B features a nominal context window of 260,000 tokens, the system actively manages the active context to stay under 100,000 tokens, where agent performance otherwise begins to decline.

The Windows release introduces local Model Context Protocol (MCP) servers, enabling the agent to interact with desktop applications and local files securely. A built-in scheduler allows users to run recurring tasks automatically while away from their computers. For tasks that exceed local reasoning capabilities, the system can escalate requests to cloud models. However, a built-in personally identifiable information (PII) classifier filters sensitive data, and the orchestrator requires explicit user permission before sending any context to the cloud.

For developers and enterprise users, this architecture shifts the cost of high-volume agent workflows from per-token API charges to local hardware depreciation and electricity. Because agent loops often require repeated planning and tool execution, sometimes accumulating hundreds of thousands of tokens per task, running them locally consumes zero account credits. This setup is particularly beneficial for privacy-sensitive operations like repository-wide code migrations or spreadsheet reconciliation.

This is our own summary of reporting by AlphaSignal

More in Agents