Start in the Cloud, Hand Sensitive Steps to a Local Model: Perplexity Built a Hybrid Inference Orchestrator on Mac

Perplexity shipped Hybrid Compute on Mac on September 1: within a single task, the hardest reasoning, search and planning stay in the cloud, and steps involving sensitive information are handed mid-task to a local model running on Apple silicon — without the user switching modes, restarting, or managing the handoff. The decision is made by a locally running Privacy Gate that detects names, addresses, account numbers and secrets, then applies one of four outcomes: keep local, mask sensitive spans, refuse, or ask for consent. Masked values are swapped for stand-ins on the way out and restored when the cloud answer returns. It launches with three local models: Gemma 4 E4B, Qwen3.6 35B-A3B, and a version Perplexity post-trained itself. Available to Pro, Max and Enterprise subscribers on Apple silicon Macs running macOS 15+ with at least 24GB of unified memory, 32GB recommended.

The Direction Is Cloud-Down, the Inverse of Last Week's Release

One thing is easy to mix up: what Perplexity shipped on NVIDIA DGX Spark the week before was local compute — starting on the user's hardware and orchestrating up to the cloud when needed. Hybrid Compute on Mac runs the other way. The task begins in the cloud, keeps the heaviest reasoning, search and planning there, and hands a step down to the local model only when it involves sensitive information. The design point is switching mid-task: the user does not change modes, restart the work, or manage the handoff. A side effect is cost — local inference handles the lighter work without consuming Computer credits, and the associated data stays on the Mac.

The Privacy Gate Is the Core of This

What determines whether this approach holds up is not how strong the local model is but how accurate the classifier is. Perplexity trained a Privacy Gate that runs locally to detect sensitive information such as names, addresses, account numbers and secrets, and applies one of four outcomes when it flags something: keep local, mask sensitive spans, refuse, or ask for consent — with credentials, payment card numbers and government IDs getting the strictest handling. Masked values are replaced with stand-ins on the way out and restored when the cloud answer comes back. Notably, the classifier is open source: PII-Tracer is a 0.6B bidirectional encoder adapted from a Qwen3 backbone. For teams building similar privacy routing, that is more useful than the product itself — it turns "what counts as sensitive" into a separate component that can be inspected and swapped, rather than a set of rules buried in a prompt.

Requirements and Boundaries

You need a Pro, Max or Enterprise subscription, any Apple silicon Mac on macOS 15 or later, and at least 24GB of unified memory, with 32GB recommended. No Ollama or manual runtime setup is required. Once a task runs, the app displays local CPU, GPU and memory utilization along with tokens consumed. Three local models ship at launch: Gemma 4 E4B, Qwen3.6 35B-A3B, and a third that is Perplexity's own post-trained version of Qwen3.6 35B, with more planned. The product page's setup flow points to a one-click download of PPLX Qwen 3.8 27B, and Perplexity's Hugging Face organization carries matching builds. The practical implication: this is one of the few approaches that turns "which data never leaves this machine" into a visible in-product mechanism rather than a privacy promise. If you evaluate it, the thing to test hardest is the Privacy Gate's false negatives and false positives — a miss means sensitive content goes to the cloud anyway, while a false positive pushes a step that belonged in the cloud onto the local model, directly affecting output quality. Perplexity has not published rates for either.

via: Perplexity's blog, Perplexity product page, MarkTechPost, The New Stack