Perplexity launched Hybrid Compute for its Computer agent on Apple silicon Macs on September 1, 2026, splitting a single agentic task between frontier models in the cloud and compact models running on the machine so confidential files never have to leave the device. An on-device classifier reads each task before anything is sent, and the company open-sourced that classifier alongside the release. It is one of the first mainstream agent products to make local-versus-cloud routing an automatic, privacy-gated decision rather than a manual toggle.
What Hybrid Compute Actually Does
Perplexity's Computer agent runs multi-step tasks: it reads files, fills forms, browses, and chains actions together. Until now every step of that work went to the cloud. Hybrid Compute changes the default so that the heavy reasoning stays in the cloud while any step that touches sensitive material gets handed to a local model on the Mac. The Hybrid Compute announcement frames this as keeping tax returns, bloodwork, and litigation files on the device by design, not by policy.
The feature is live now for Pro, Max, and Enterprise subscribers, as 9to5Mac reported. Setup is one click with no Terminal commands or runtime installs, and an iPhone or iPad can offload its sensitive steps to a Mac running on the same account. That last detail matters: the phone does not need to run a local model itself, it just routes the private part of a task to a machine that can.
The Privacy Gate: An Open-Sourced PII Classifier
The core of the system is a small model Perplexity calls PII-Tracer. It is a 0.6-billion-parameter bidirectional encoder adapted from a Qwen3 backbone, with a tagging head that emits 37 labels across nine categories of personally identifiable information and a second head that predicts whether a conversation contains sensitive material at all. Perplexity says it trained the classifier with its Secure Intelligence Institute and open-sourced it, which is unusual for a component that sits this close to a paid product.
Before any content from a protected file reaches the cloud, the gate applies one of four outcomes: keep the step local, mask the sensitive spans and continue, refuse the action, or ask the user for consent. When masking is used, names, addresses, and account numbers are swapped for stand-ins on the way out and restored when the answer comes back, so the cloud model never sees the real values but the final result still reads correctly.
Cloud-to-Local Orchestration Compared
Most on-device AI ships as an all-or-nothing choice: either a task runs entirely in a local model, with the quality ceiling that implies, or it runs entirely in the cloud. Hybrid Compute is closer to routing than to a fallback. The cloud model stays the orchestrator and delegates only the steps that trip the privacy gate to a local model. At launch the local options are Google's Gemma 4 E4B, Alibaba's Qwen3.6 35B-A3B, and a Perplexity post-trained Qwen variant.
| Approach | Where reasoning runs | Privacy model | Quality tradeoff |
|---|---|---|---|
| Cloud-only agent (default before this) | Cloud | All data sent to cloud | Highest capability, no local privacy |
| Local-only assistant | Device | Nothing leaves device | Private but capped by local model size |
| Perplexity Hybrid Compute | Cloud, with sensitive steps local | On-device gate masks or keeps PII local | Cloud capability with local privacy on sensitive spans |
How to Set It Up and Use It
The requirements are specific. Hybrid Compute runs on Apple silicon on macOS 15 or later, with a 24 GB unified-memory minimum and 32 GB recommended for best results. If your Mac meets that bar, the workflow is short:
1. Update the Perplexity Mac app and confirm you are on a Pro, Max, or Enterprise plan, which are the tiers that receive the feature.
2. Enable Hybrid Compute in the Computer agent settings and pick a local model. The first choice downloads once and then runs offline for the local steps.
3. Point the agent at a real task that involves a private document, such as summarizing a medical letter or extracting figures from a tax PDF, and watch the gate decide which steps stay on device.
4. Review the gate's decisions. When it asks for consent, that is the moment to confirm what leaves the machine. For creators handling client contracts or unreleased work, this is the setting that turns an agent from a liability into a usable tool.
Why It Matters for Creators and Builders
For anyone building on top of AI agents, the interesting part is not the Mac app but the pattern. A cloud orchestrator that can delegate to a gated local model is a template other agent platforms will copy, because it resolves the tension that has kept agents out of regulated and confidential work. Creators sitting on unreleased scripts, client NDAs, or personal financial records have had to choose between capable cloud tools and private local ones. Hybrid Compute is an argument that you do not have to.
The open-sourced classifier also lowers the bar for builders who want the same guarantee in their own stacks. A 0.6B PII detector that runs locally is small enough to embed in a desktop app or a browser extension, which means the privacy-gate idea is not locked inside Perplexity. Expect the same split-execution approach to show up in coding agents and creative pipelines that already lean on local open models.
Frequently asked questions
What is Perplexity Hybrid Compute?
It is a feature for Perplexity's Computer agent on Apple silicon Macs that splits a single task between cloud models and local models, keeping sensitive data on the device through an on-device privacy gate. It launched on September 1, 2026.
Which Macs can run it?
Apple silicon Macs on macOS 15 or later, with at least 24 GB of unified memory and 32 GB recommended. Intel Macs are not supported. You can also run a nearby Mac as the local endpoint for tasks started on an iPhone or iPad.
Which local models does it use?
At launch the local options are Gemma 4 E4B, Qwen3.6 35B-A3B, and a Perplexity post-trained Qwen variant. The cloud still handles the heavy reasoning; the local model only runs the steps the privacy gate keeps on device.
Is the privacy classifier really open source?
Yes. Perplexity open-sourced PII-Tracer, a 0.6B encoder that tags 37 label types across nine PII categories, and says it was trained with the company's Secure Intelligence Institute. That makes the gate auditable and reusable outside Perplexity.
Do I need a paid plan?
Yes. Hybrid Compute is available to Pro, Max, and Enterprise subscribers at launch. There is no Terminal setup or manual runtime install; enabling it is a one-click step in the app.
What happens to my data when a step goes to the cloud?
The gate can keep the step local, mask sensitive spans with stand-ins, refuse the action, or ask for your consent. When masking is used, real names, addresses, and account numbers are replaced before transmission and restored in the returned answer, so the cloud model never sees the originals.
