Perplexity's Hybrid Compute Keeps Your Private Files Off the Cloud, Explained
On September 1, 2026, Perplexity rolled out a new "Hybrid Compute" mode for its Mac app that splits every AI task between the cloud and your own machine -- and open-sourced the exact system it uses to decide which parts of your data are too sensitive to leave your computer. Here's how it actually works, what hardware you need, and who can use it today.
⚡ Quick facts
- Launched: September 1, 2026, inside the Perplexity Mac app
- What it does: splits Perplexity Computer tasks between cloud models and a local model on your Mac, so sensitive steps never leave the device
- Requirements: Apple Silicon Mac, macOS 15+, minimum 24GB unified memory (32GB recommended)
- Who gets it: Perplexity Pro, Max, and Enterprise subscribers -- local processing uses no cloud credits
What Hybrid Compute actually does
Perplexity Computer -- the company's agent that browses the web, edits local files, and operates your Mac on your behalf -- normally routes every step of a task to the cloud. Hybrid Compute changes that for tasks that touch anything private. According to Perplexity's own announcement (confirmed via its official X post), a task can now start in the cloud -- where frontier models handle reasoning, web search, and planning -- and then hand off specific steps to a model running entirely on your Mac whenever the task involves a local file or sensitive information. The two halves work as one continuous task; you don't have to manually copy anything or run a second prompt.
The "Privacy Gate": what decides what leaves your Mac
The feature is built around what Perplexity calls a Privacy Gate -- an on-device classifier that scans for personally identifiable details like names, addresses, and account numbers before any of that content is sent to the cloud. Depending on what it finds, the gate can mask the sensitive parts and restore them once the cloud's answer comes back, keep the entire step local, refuse to send the request at all, or simply ask you for consent first.
Perplexity has also open-sourced the classifier itself. CEO Aravind Srinivas said the company is "open-sourcing the PII classifier that we use for deciding when to send the workload to the local model" -- letting outside researchers inspect exactly how the gate makes that call, rather than asking users to simply trust it.
What you need to use it
Hybrid Compute only runs on Apple Silicon Macs with macOS 15 or later and at least 24GB of unified memory (Perplexity recommends 32GB for smoother performance). It ships with three local models to choose from: Google's Gemma 4 E4B, Alibaba's Qwen3.6 35B-A3B, and a Perplexity model specifically post-trained for Perplexity Computer. Setup is three steps -- install the latest Perplexity Mac app, download a local model in a single click (no manual runtime like Ollama required), then open the model selector, choose Hybrid, and pick which local and cloud models to pair.
The feature is live now for Perplexity Pro, Max, and Enterprise subscribers. Work handled by the local model doesn't draw down any cloud credits and doesn't need an API key, since it runs entirely on the Mac's own hardware once the model is downloaded. Enterprise admins additionally get organization-wide controls to set rules for what must stay local, what gets masked, and what needs approval before it can go to the cloud.
Why this matters
Most AI assistants that operate your computer -- reading files, filling forms, browsing on your behalf -- send everything they see to a cloud model by default, which is a real concern once an agent starts touching tax documents, medical records, or business files. Hybrid Compute is one of the more concrete answers yet from a major AI company to that specific worry: instead of asking users to simply trust that sensitive data is handled responsibly, Perplexity is publishing the actual code that decides what counts as sensitive in the first place.
Local CoreML vs Cloud Orchestration Routing Matrix
Perplexity Hybrid Compute for macOS implements a deterministic routing layer between Apple's CoreML neural engine and Perplexity's cloud search infrastructure. Queries classified as privacy-sensitive or requiring verified citation retrieval are sent over encrypted tunnels to Perplexity servers, while conversational and summarization requests that do not involve external data access remain entirely local.
Frequently asked questions
What is Perplexity's Hybrid Compute?
Hybrid Compute is a mode in the Perplexity Mac app, launched September 1, 2026, that splits an AI task between the cloud and your own Mac. Cloud models handle reasoning, web search, and planning, while a model running locally on your machine handles steps that touch private files or sensitive information.
How does the Privacy Gate decide what stays on my Mac?
An on-device classifier scans for sensitive details like names, addresses, and account numbers before anything is sent to the cloud. Depending on what it finds, the gate can mask the sensitive parts, keep the whole task local, refuse to send it, or ask you for consent first. Perplexity has open-sourced this classifier so outside researchers can inspect how it makes that call.
What do I need to use Hybrid Compute?
An Apple Silicon Mac running macOS 15 or later, at least 24GB of unified memory (32GB recommended), and a Perplexity Pro, Max, or Enterprise subscription. The local models -- Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity model tuned for Perplexity Computer -- download in one click with no separate setup.
Does using the local model cost extra?
No. Work handled by the local model doesn't use any cloud credits and doesn't require an API key -- it runs entirely on your Mac's own hardware once the model is downloaded.