Perplexity has launched Hybrid Compute, a feature for its Mac app that splits a single AI agent task between the cloud and a small model running locally on the device, so sensitive information never has to leave the machine. The company introduced the feature on September 1, rolling it out to Pro, Max, and Enterprise subscribers running the Perplexity Computer agent on Apple silicon Macs.
How it works
Perplexity’s Computer agent normally starts a task in the cloud, where frontier models handle web search, planning, and reasoning. With Hybrid Compute turned on, an on-device classifier the company calls a Privacy Gate screens each step before it leaves the Mac. When a step touches something sensitive — a password, a government ID, a payment card number — the task hands off to a compact local model instead, then merges the result back into the cloud-run workflow without restarting or losing context.
The screening itself runs on a small model Perplexity built for the job: a roughly 0.6-billion-parameter classifier built on a Qwen3 backbone, tuned to recognize dozens of categories of personal data. For the local processing step, users can pick from several open models, including Gemma 4 E4B and Qwen3.6 35B-A3B, or let a Perplexity-tuned model download automatically the first time the feature runs — no separate runtime or API key required.
Requirements and cost
Hybrid Compute needs an Apple silicon Mac running macOS 15 or later, with at least 24GB of unified memory (32GB recommended). Perplexity said it does not charge for the local portion of a task — billing applies only to cloud model usage. Enterprise customers additionally get organization-wide rules for what stays on-device and audit logging for compliance teams.
Why it matters
The launch follows a broader industry push toward on-device AI as companies look to reassure users and enterprises handling regulated or confidential data. OpenAI took a similar approach last month with its own private, on-device safety screening, and rivals building computer-use agents that act independently on a user’s desktop face the same pressure to prove they won’t leak what they see on screen.