Perplexity Ships Hybrid AI That Keeps Sensitive Work On Apple Silicon Macs

Perplexity's Computer agent now hands sensitive steps to a locally running model on the user's Mac while keeping heavy reasoning in the cloud.

Perplexity Ships Hybrid AI That Keeps Sensitive Work On Apple Silicon Macs

Perplexity has launched a hybrid-compute capability inside its Computer agent that dynamically routes work between cloud frontier models and open-weight models running on the user's Apple silicon Mac. The transition, the company says, happens without restarting the workflow or losing agent context.

How The Routing Works

A task typically begins in the cloud, where a large model plans and reasons. When any step touches confidential material — a contract, an internal codebase, personal financial data — the runtime hands that step to a locally hosted open-weight model on the Mac, executes it against on-device files, and stitches the result back into the same agent trace.

The Enterprise Objection Being Answered

Data exposure has been the single biggest enterprise objection to autonomous agents. CIOs are comfortable letting frontier models plan; they are much less comfortable letting them read a payroll spreadsheet, a customer database or a legal draft. Hybrid inference makes it possible to keep the powerful reasoning in the cloud without pushing sensitive documents outside the network.

AI agent workflow on a laptop

Why Apple Silicon Specifically

Apple silicon is the obvious runtime because M4- and M5-class chips already run 8B–70B parameter open-weight models at usable speed with unified memory. The move also feeds a hardware narrative: the more capable local models become, the more Apple's Neural Engine roadmap becomes an infrastructure story, not just a consumer story.

Not Just Privacy — Also Latency And Cost

Even outside privacy scenarios, local execution helps with latency-sensitive UI steps and cuts cloud inference cost. For agents that run for hours doing rote work, sending only the difficult reasoning to the cloud can materially change the per-run bill.

Context

The launch coincides with Anthropic's model hardware standard for lab robots, Apple's M6 2nm rollout and Anthropic's $35B Lambda cloud deal — snapshots of an industry now designing every layer of the stack around agentic workloads.

Reporting based on coverage from VentureBeat, The Verge and Tech Startups.

Category: Edge Computing

Tags: Robotics Edge Computing AI Physical AI AI Agents AI Infrastructure

Related Articles