Perplexity's Hybrid Local-Cloud Answer to Private AI

Updated June 11, 2026 · ~2 min read · Ilura Technology

Source: VentureBeat

What was announced?

According to reporting by VentureBeat from Computex 2026, Perplexity used Intel’s keynote to demo a hybrid local-server inference system. The orchestrator splits a single request between the device and the cloud. Pieces it judges sensitive stay on the machine and run locally. The rest goes to servers. Before anything leaves the device, the system reportedly asks the user for permission. Per the same report, the capability ships in Perplexity Computer in July.

The details come from press coverage rather than a first-party technical post, so specifics may shift by release.

What does it mean for personal, on-device writing?

This is the most credible competing privacy story so far. It is also a quiet admission. A company does not build consent prompts and local routing unless it accepts that pure-cloud assistants have a trust problem.

The hybrid model’s strength is reach. When the device cannot handle a request, a data-center model can. Its weakness is the seam. Request by request, software must decide which parts of your words count as sensitive. That decision is a judgment call made by a classifier, and classifiers make mistakes in both directions. And each permission dialog is one more prompt a busy person clicks through without reading. Consent that arrives forty times a day stops working as consent.

Hybrid routing makes privacy a runtime decision. The alternative is to make it an architecture decision, settled once, before any request exists.

Where does ILURA stand?

This deserves an honest comparison, because it is the strongest rival framing of “private AI” yet.

ILURA’s model is simpler: nothing leaves the device at all. Drafts run on Apple’s on-device foundation models through Apple Intelligence. There is no account, no routing layer to audit, no classifier deciding what is sensitive, and no permission dialog to misread. The App Store privacy label reads Data Not Collected because there is nothing to collect.

The trade is real and worth stating. A fully local app cannot borrow data-center power for heavy tasks; Perplexity’s hybrid can. But ILURA is not trying to be a research engine. It drafts personal correspondence — the follow-up, the brief, the difficult no — guided by role playbooks you train by correcting. On-device models are already enough for that job. For writing that carries your name, the simplest privacy story is the one with no moving parts.