Perplexity Hybrid Compute runs sensitive steps on a local Mac model
A cloud agent orchestrates; an on-device classifier masks your PII and hands private steps to a downloadable Qwen model. Pro and up, no cloud credits.
Copy markdown
Cloud starts the task, your Mac finishes the private part
Trigger a task from your iPhone or desktop and Perplexity runs the reasoning, planning, and web search in the cloud — but any step touching your local files or sensitive data executes on-device on your Mac. It's a split-brain agent, not a fully-local one.
An on-device classifier masks PII before anything leaves
A local classifier scans for personal or sensitive details, swaps them for stand-ins before cloud transmission, then restores them on completion — and can also mask, refuse an action outright, or pause to ask for consent. The privacy gate is the actual product here.
What runs locally: a one-click Qwen, 24GB Mac floor
The default on-device model is Perplexity's tuned PPLX Qwen 3.8 27B (Gemma 4 E4B and Qwen3.6 35B-A3B are also selectable), downloaded in one click with no API keys or runtime config. It needs an Apple-silicon Mac on macOS 15+ with at least 24GB unified memory (32GB recommended); 8GB and 16GB machines can't run it.
Local runs cost zero cloud credits — but it's Pro-and-up
Work handled on-device uses no cloud credits and needs no API key, so private tasks don't burn your token budget. It's gated to Perplexity Pro, Max, and Enterprise subscribers and wired into the Computer agent.
Elsewhere: Anthropic will keep your Claude logs in your own cloud
Anthropic's new Enterprise Frontier Safeguards lets monitoring activity data live in your own S3, Azure Blob, or GCS bucket under your keys, access policies, and audit logging — free, and slated to cover Claude Code, the Claude Platform, Bedrock, and Microsoft Foundry. Broad availability is targeted for later this fall.