Perplexity's Hybrid Compute: How It Keeps Private Files Off the Cloud
Perplexity's hybrid compute uses a local Privacy Gate to keep sensitive files on your Mac while sending other tasks to cloud models.
What is Perplexity’s hybrid compute system?
Hybrid compute is a feature in Perplexity’s Mac app that splits a single task between an on-device model and Perplexity’s cloud models. Instead of choosing local-only or cloud-only processing, the app decides which pieces of a request can be handled safely on your Mac and which pieces need the broader capability of a cloud model, then stitches the results back together. The goal is to let you hand over tasks involving sensitive files, like bank statements or medical records, without those files actually leaving your machine.
TL;DR
- Hybrid compute routes parts of a task to a local on-device model and parts to cloud models automatically, instead of forcing an all-or-nothing choice.
- Privacy Gate is the built-in checkpoint that scans files and content for personal information before anything gets sent off the Mac.
- A small open-source model called PII Tracer, published by Perplexity on Hugging Face, does the actual detection of personal data in documents and images.
- You can pick from a few local orchestrator models, including Gemma 3 4B, Qwen 3.6 35B, and a custom Perplexity model, and the options available depend on your Mac’s RAM.
- In testing, cloud models handled web research and planning (like comparing internet service rates), while the local model handled document parsing and OCR on sensitive files without uploading them.
- When private data is detected, the app pauses and asks whether to process locally or share with the cloud, giving the user a visible decision point rather than a silent default.
- Lower-RAM Macs get fewer local model choices, which means the privacy and capability tradeoffs shift depending on your hardware specs.
Everyone else built a construction worker.
We built the contractor.
One file at a time.
UI, API, database, deploy.
Why does local versus cloud AI matter for privacy?
Cloud AI models are generally more capable because they run on much larger infrastructure than a laptop can offer. But using them usually means sending your data, files, prompts, context, to a remote server. For everyday tasks like drafting an email or debugging code, that’s an easy tradeoff most people accept. For financial records, medical documents, or anything with identifying numbers attached (account numbers, ticket IDs, medical procedure details), that tradeoff gets uncomfortable.
Local, on-device models solve the privacy problem by never sending data anywhere, but they’re constrained by the hardware they run on. A model small enough to run on a 16 GB laptop is not going to match the reasoning or research capability of a frontier cloud model. Hybrid compute is Perplexity’s attempt to close that gap by using each type of model for what it’s actually good at: cloud models for synthesis, web lookups, and complex reasoning, local models for reading and filtering the parts of a task that contain private data.
How does Privacy Gate decide what stays local?
Privacy Gate acts as a filter that runs before any content is routed to the cloud. When you attach files or point the app at a folder, the local model reads through that content and checks it against Perplexity’s PII Tracer, a small model built specifically to detect personal information. If PII Tracer flags something, the app stops and surfaces a review prompt asking how to handle it, with the Mac-only option recommended by default.
In practice, this means a request like “review six months of bank statements and flag charges that don’t match my service agreement” gets split automatically. The local model parses the statements, bills, and agreement on-device, strips out account numbers and other identifying details, and only passes the non-sensitive, task-relevant information (like a list of charges and dates) to the cloud model for comparison and web research. The redacted content and the content that was sent can both be inspected after the fact, since the app shows a breakdown of what was filtered.
The same filtering extends to images. Photos of documents, like a parking ticket, a medical record, or a car rental receipt, get processed with on-device text recognition rather than being uploaded, as long as hybrid mode and Privacy Gate are active. Attaching an image directly in the chat window bypasses this protection, since direct attachments go straight to the cloud, so the privacy benefit depends on using the file or folder attachment path that triggers local processing first.
What local models does Perplexity support, and does hardware matter?
The local orchestrator model options scale with your Mac’s available memory. The models mentioned are Gemma 3 4B (an open-weights model, roughly 6.6 GB at 4-bit quantization, recommended for Macs with at least 16 GB of RAM), Qwen 3.6 35B (open-weights, recommended for 32 GB of RAM), and a custom Perplexity model (around 19 GB, also recommended for 32 GB of RAM).
This creates a real tiering effect. A 16 GB MacBook Air only gets access to Gemma 3 4B. A machine with 32 GB of RAM opens up all three options, including Perplexity’s own tuned model. Machines with significantly more memory don’t get additional model choices beyond that tier, since the ceiling is set by the most capable model Perplexity currently offers for local inference, not by how much spare RAM you have. The practical upshot: the privacy protections work on any supported Mac, but the quality and speed of the local reasoning step improves on machines with more memory.
Is hybrid compute worth using over a cloud-only or local-only setup?
For anyone who already uses AI tools for tasks involving financial, medical, or other sensitive documents, hybrid compute addresses a specific gap: it removes the manual effort of deciding, file by file, what’s safe to upload. Without a system like this, achieving the same privacy outcome means either avoiding AI for sensitive documents entirely or manually redacting information before pasting it into a chat window, both of which are slower and error-prone.
The tradeoffs are real, though. Local models are weaker than frontier cloud models, so tasks that are entirely local (no cloud research needed) will be capped by whatever the on-device model can do. The system also depends on PII Tracer correctly identifying what counts as private, and while the demonstrated runs caught bank account details, medical records, and ticket numbers, no detection system is guaranteed to catch every edge case. Users who want hard guarantees should still treat the review prompt as a checkpoint to actually read, not a rubber stamp.
For anyone running a lower-RAM Mac, it’s also worth knowing upfront that the model selection will be limited to Gemma 3 4B, which affects how much of the workflow can run locally before cloud assistance is needed for heavier tasks like deep web research.
Frequently Asked Questions
What is Privacy Gate in Perplexity’s Mac app?
Privacy Gate is a feature that scans content and files for personal information before sending anything to cloud models. It uses Perplexity’s open-source PII Tracer model to detect sensitive data and pauses to ask the user whether to process that content locally or share it with the cloud.
Does hybrid compute work without an internet connection?
The transcript doesn’t address fully offline use. Hybrid compute by design relies on a cloud component for tasks like web research, so a cloud-dependent step will need connectivity even though the private data itself stays local.
Which Macs can run Perplexity’s local models?
Model availability depends on RAM. Gemma 3 4B needs at least 16 GB of RAM, while Qwen 3.6 35B and Perplexity’s own model need at least 32 GB. Macs with 32 GB or more get access to all three local model options.
Does attaching a photo always keep it private?
No. Attaching an image directly in the chat sends it straight to the cloud. To get on-device processing and text recognition, files need to be added through the folder or file attachment flow with hybrid mode and Privacy Gate enabled.
Is PII Tracer open source?
Yes. PII Tracer is published by Perplexity AI on Hugging Face as an open-source model, which means its detection behavior can be inspected independently rather than taken purely on faith.