Perplexity: Local AI Agent, Linux Then Windows in September

Perplexity launches Portable Computer, a local version of its AI agent reserved for paying subscribers and currently limited to Linux machines. Windows support is expected in September, with the requirement of an NVIDIA RTX GPU with at least 24 GB of VRAM. The agent runs models on the device, allowing the user to decide on cloud connectivity, and local processing incurs no inference fees according to Perplexity.

Paid Access, Linux First, and a Required RTX GPU

Portable Computer is not available for free accounts: it is reserved for Pro, Max, Enterprise Pro, and Enterprise Max offerings. For now, the agent is only accessible on Linux. Perplexity plans to support Windows in September. On the hardware side, under Linux, a Nvidia DGX Spark or another computer equipped with an Nvidia RTX GPU is required, and it must run on Nvidia DGX OS or Ubuntu on ARM or x64. Under Windows, the required configuration includes an NVIDIA RTX GPU with at least 24 GB of VRAM, a threshold that many desktop computers may not meet.

On-Device Processing and Optional Cloud Service Connectivity

The agent offers the usual commands, features, and AI tools and can perform tasks with minimal intervention. Models, files, and jobs remain on the user's machine. The user decides if and when the agent connects to the Internet. Connectors allow integration of content hosted on Google Drive, Gmail, Slack, or GitHub.

Performance, Security, and Costs: The Case for Local

Running models directly on the device allows for faster responses than with online services, while ensuring that operations occur in an isolated environment. When calculations are performed in the cloud, this involves the consumption of credits and tokens, whereas local usage requires no billing credits; if a task is partially processed in the cloud, only that fraction is accounted for. Perplexity describes Portable Computer as a primarily local agent, specifying that sensitive information remains on the device unless otherwise authorized, that the use of local models incurs no inference costs, and that the system is designed to ensure privacy and limit expenses.

Qwen 3.8 27B, PPLX 27B, and an Announced Nvidia Model

Among the available models is Qwen 3.8 27B. A post-trained variant, PPLX 27B, is also offered. A post-trained model benefits from additional fine-tuning beyond its initial instructions, and Perplexity states that it post-trains all models to improve accuracy and efficiency. Nvidia plans to launch Nemotron 3.5 Lightning, designed for heavy and prolonged workloads. The user will be able to switch between models based on the tasks assigned to the agent.