Perplexity Hybrid Compute: Which Mac Runs Which Local Model?
Perplexity's hybrid compute keeps private files on your Mac. Here's which local models (Gemma, Qwen, Perplexity's own) run on 16GB, 32GB, and 128GB Macs.
What is Perplexity’s hybrid compute?
Perplexity’s hybrid compute is a system built into its Mac app that splits a single task between a cloud model and a small local model running on your own hardware. A feature called privacy gate inspects files, photos, and documents before anything leaves your machine, stripping out personal details (names, account numbers, medical info) and sending only the non-sensitive parts to the cloud for the heavier reasoning and web research. The sensitive stuff gets processed locally by an on-device model instead, and never gets uploaded.
TL;DR
- Privacy gate is the filtering layer that scans attached files and photos, redacts personal information, and decides what’s safe to send to cloud models versus what has to stay local.
- Three local orchestrator models are currently offered: Gemma 4 E4B (4-bit, 6.6GB, needs 16GB of RAM), Qwen 3.6 35B (needs 32GB of RAM), and Perplexity’s own fine-tuned model (19GB, also needs 32GB of RAM).
- RAM tier determines model choice: a 16GB MacBook Air only opens up Gemma 4 E4B, while Macs with 32GB or more (a Mac Mini or MacBook Pro) open up all three options including Qwen 3.6 and Perplexity’s custom model.
- The small local model called PII Tracer, an open-source model Perplexity published on Hugging Face, is the piece actually doing the personal-information detection before the privacy gate decision is shown to the user.
- The workflow is interactive, not silent: when privacy gate finds sensitive content, it stops and asks the user to choose “process on my Mac” or “share,” rather than making that call automatically.
- Image attachments bypass local filtering if dropped straight into the chat window, meaning photos sent that way go to the cloud immediately. Using the “your Mac” folder reference instead keeps them local and triggers on-device text recognition.
- Hybrid mode worked across a real test case involving bank statements, an ISP contract, and bills, correctly matching charges, flagging an $88 overcharge, and finding cheaper internet deals through live web search while keeping financial documents off the cloud.
Remy doesn't write the code. It manages the agents who do.
Remy runs the project. The specialists do the work. You work with the PM, not the implementers.
How does hybrid compute decide what stays local?
The decision runs through two layers. First, privacy gate scans whatever you attach (folders, PDFs, images) using a lightweight detection model, Perplexity’s PII Tracer, to flag anything that looks like personal or financial data: account numbers, addresses, medical record details, ticket numbers. Second, once something sensitive is found, the app pauses and surfaces an explicit choice: process the flagged content on your Mac, or share it with the cloud model.
In practice, this means the cloud-based “planner” model handles orchestration and anything requiring live web search, like comparing your ISP contract’s rate against current competitor pricing. The local model handles direct interaction with your private files: reading a bank statement, matching line items against a contract, running OCR on a photographed medical document. Only the redacted, non-sensitive summary gets passed upward to the cloud model for the parts of the task that need external data.
One wrinkle worth knowing: if you attach a photo directly into the chat composer rather than pointing the assistant at a folder on your Mac, the image uploads immediately, before privacy gate gets a chance to intervene. The safer pattern demonstrated was referencing files by location (“look at the file on my desktop”) so the local model reads and filters them first.
Which Mac specs do you need for each model?
Hardware RAM is the gating factor, and Perplexity’s app enforces it directly in the model picker rather than letting you attempt to load something your machine can’t handle.
- 16GB of RAM (tested on an M4 MacBook Air): only Gemma 4 E4B is offered. It’s a 4-bit quantized open-weights model at roughly 6.6GB, sized specifically to fit in a 16GB unified memory budget alongside the OS and the app itself.
- 32GB of RAM (tested on an M6 Mac Mini): this opens up all three local orchestrator options, Gemma 4 E4B, Qwen 3.6 35B, and Perplexity’s own 19GB fine-tuned model. The Mac Mini sat right at the 32GB threshold and still had access to the full set.
- 128GB of RAM (tested on a MacBook Pro): comfortably runs every available local model with no constraints, though the app doesn’t currently offer anything larger than the 32GB-tier models, so the extra memory headroom doesn’t open up additional choices beyond what a 32GB machine already gets.
The practical takeaway: 32GB is the meaningful upgrade point. Going from 16GB to 32GB expands your choice of local orchestrator from one model to three. Going beyond 32GB, at least with the current model lineup, doesn’t add new tiers.
Is Gemma 4 E4B good enough on a 16GB Mac?
Other agents start typing. Remy starts asking.
Scoping, trade-offs, edge cases — the real work. Before a line of code.
For the kind of filtering and light local reasoning hybrid compute asks of it, yes, based on the demonstrated workflow. Gemma 4 E4B’s job in this system isn’t to do heavy-duty analysis. It’s to read your private documents, extract what’s needed, strip out personal identifiers, and hand off a clean summary to the cloud model for the parts of the task that benefit from bigger context and live search. That’s a narrower, more tractable job than general-purpose chat, which is why a 4-bit, 6.6GB model can plausibly handle it on 16GB of unified memory.
The tradeoff is choice. Owners of 16GB machines get exactly one local model option. If Perplexity adds more models to the lineup in the future, there’s no guarantee they’ll fit in that same memory envelope, so Gemma 4 E4B may remain the only path for base-tier Mac hardware for a while.
Does hybrid compute actually keep data private?
In the demonstrated test, yes, with the system behaving as described. A folder of fabricated financial documents (bank statements, an ISP service agreement, and bills) were attached to a hybrid-mode session, and privacy gate intercepted the request before processing, asking explicitly whether to handle the content on-device or share it with the cloud. Choosing local processing kept the financial documents on the Mac while the cloud model still performed the web research needed to compare rates and flag a billing discrepancy.
Photos of a car rental receipt, a medical document, and a parking ticket were tested separately. Each one triggered the same privacy gate prompt, and in each case the on-device model ran text recognition locally rather than uploading the image. For the medical document, the system answered a general question about the procedure using web search while explicitly keeping the document’s private details off the cloud. For the parking ticket, it refused to send the ticket number anywhere and instead told the user to look up the number manually against public records, since the number itself counts as identifying information.
None of this is a formal security audit, just a hands-on run-through. But the visible behavior (the GPU activity during local inference, the redaction log showing exactly what was filtered versus what was sent) suggests the privacy gate is doing real work, not just a cosmetic toggle.
Frequently Asked Questions
What local models work with Perplexity’s hybrid compute?
Three options exist at the time of testing: Gemma 4 E4B (open weights, 4-bit, 6.6GB), Qwen 3.6 35B (open weights), and a 19GB custom model built by Perplexity. All three require macOS and Apple Silicon memory to run on-device.
How much RAM do I need to run Perplexity’s local orchestrator models?
Gemma 4 E4B needs at least 16GB of RAM. Qwen 3.6 35B and Perplexity’s own 19GB model both need at least 32GB of RAM. Machines with more than 32GB don’t currently open up any additional models beyond that tier.
What is privacy gate in Perplexity’s app?
Privacy gate is the filtering feature that scans files and images for personal or sensitive information before any cloud processing happens. It uses a small open-source model called PII Tracer to detect personal data, then prompts the user to choose whether to process that content locally or share it with cloud models.
Can I force an image to stay local in Perplexity’s hybrid mode?
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
Dropping an image directly into the chat composer uploads it to the cloud immediately, bypassing local filtering. To keep an image local, reference its location on your Mac instead (for example, asking the assistant to look at a file on the desktop), which triggers on-device text recognition rather than a direct upload.
Does using local models in hybrid mode cost credits?
No. Based on the demonstrated setup, the local model handling private files on-device does not consume cloud credits. Credits are tied to the cloud-side planning and web research portions of a hybrid task.