Perplexity, the AI search company, on Tuesday launched Portable Computer, an agent that runs its model, harness and conversation history on a user's own hardware instead of a remote cluster.
The whole setup runs locally by default, and when a user needs a stronger AI model, the system seeks the user's approval before escalating to the cloud, according to a blog post from the company's engineering team.
"Sensitive data therefore never leaves the device without permission, and local models carry no inference fee," according to the blog post. "The system is private and cost-effective by construction."
The design reads like a list of virtues decentralized AI projects have been circulating for years: open-weight models, computation on the device and privacy enforced by architecture rather than by a vendor's promise. That argument has been made repeatedly from the crypto side, including in a CoinDesk opinion piece in February.
Perplexity said that the new agent harness runs the open-source AI model Qwen 3.8 27B on an Nvidia DGX Spark, a $4,699 miniature desktop computer optimized for artificial-intelligence processing.
Using the local AI model with escalation to the more powerful Claude Opus 5 resulted in a Terminal Bench 2.1 score of 73.0% at roughly $0.415 per task, versus a score of 82.4% and $0.65 on Opus 5 alone, according to the post.
HOW AI WAS USED IN THE PRODUCTION OF THIS PIECE: I drafted this story on Claude Desktop using a skill designed to produce news-story drafts based on blog posts. I then filed the story to DeAI News on Distro using our Distro Publisher MCP connector. I used DistroVerse's built-in (non-AI) editing interface to do a thorough edit, and I checked the facts prior to publication.