update · TowCue Editorial Team
Perplexity Hybrid Compute on Mac: what changes when AI can keep sensitive work local
Perplexity Computer now splits work between cloud models and a local model on Mac, keeping sensitive files and personal data on-device when possible. Here is what that changes for privacy-sensitive AI workflows.
Quick answer
Perplexity announced Hybrid Compute on Mac on September 1, 2026. The idea is simple but important: Perplexity Computer can split one task between frontier models in the cloud and a local model running on the Mac.
Cloud models handle work such as research, reasoning and planning. Local inference can handle private files, app context and sensitive information that should not automatically leave the device. Perplexity says this routing model is designed to keep sensitive content local unless escalation to the cloud is needed and permitted.
The TowCue takeaway is not "local AI replaces the cloud." It is the opposite: hybrid AI is becoming a practical way to use cloud intelligence without treating every piece of context as cloud-safe by default.
For broader context, read TowCue's guide to choosing an AI tool for work, AI agent best-practices guide, and research stack blueprint.
What Perplexity changed
Perplexity Computer's Hybrid Compute architecture divides work according to where it should run.
The cloud remains useful for tasks that benefit from larger models, web access and heavier reasoning. The local side handles information that is better kept close to the user, including files and personal context on the Mac.
Perplexity's September 1 research also describes a local inference engine called Lily, built specifically for Apple silicon and Qwen3.6-35B-A3B. In Perplexity's own benchmark on an M5 Max MacBook Pro with 128 GB of unified memory, Lily averaged 1.23× MLX-LM's prefill throughput and 1.35× its decode throughput across the tested prompt and context lengths.
Those numbers are vendor-run measurements on one specific hardware and model configuration, so they should not be treated as a universal Mac performance claim. What matters more for most users is that Perplexity is investing in making the local side fast enough that hybrid routing does not feel like a slow privacy fallback.
Why the privacy boundary matters
Most AI assistants become more useful when they can see more context: documents, email, calendars, browser history, files, conversations and internal data.
That creates a simple tension:
- more context can improve usefulness;
- more context can also increase privacy exposure;
- cloud-only systems often force a binary choice between sharing the data or withholding it;
- hybrid systems can create a third option: process sensitive context locally, then escalate only the approved portion of the task.
Perplexity's PII-TRACE work makes the same design problem explicit. The company describes a local privacy gate that detects personally identifiable information before content leaves the device. The application can keep the relevant input local, redact detected spans, or ask for explicit approval before sending information to a cloud model.
That is a more useful mental model than a vague "private AI" label. The important question is where each step runs and what exact data crosses the boundary.
The hidden difficulty: privacy has to survive long conversations
A one-time privacy check is not enough for an agent that works across dozens of turns.
A phone number, email address or customer identifier can appear early in a conversation and then reappear much later. If a detector catches the first mention but misses the fifth, sensitive data can still leave the device.
Perplexity introduced PII-TRACE to evaluate that problem across long, multi-turn conversations. The benchmark contains 13,148 synthetic user-assistant conversations across 13 languages and 10 writing systems, with 37,431 labeled identifier mentions across nine PII types.
Perplexity pairs the benchmark with PII-Tracer, a compact 0.6B detector designed to run locally. The company says the detector provides a routing signal so the application can decide whether to keep content local, redact it or request permission before cloud escalation.
For TowCue, this is the more interesting part of the launch: hybrid AI only works if the routing layer is reliable enough to understand what should stay local.
What this means for knowledge workers
Hybrid Compute is most relevant when the task mixes private local context with information or reasoning that benefits from cloud models.
Examples include:
- summarizing a private folder, then researching public background on the web;
- reviewing internal notes, then drafting a market brief using external sources;
- finding information in local documents without uploading the entire folder;
- preparing a client or finance workflow where only selected non-sensitive context should reach the cloud;
- running an agent across local apps while keeping personal identifiers on-device.
This is different from fully local AI. Fully local systems maximize data locality but usually trade off model choice, web access or frontier capability. Hybrid Compute tries to preserve both sides.
Who should care most
Professionals working with sensitive documents
Researchers, consultants, finance teams, legal users and operators often need AI to work across documents that should not casually be copied into cloud prompts.
A local-first routing layer can reduce unnecessary data movement, but it does not eliminate the need to review retention policies, permissions and enterprise controls.
Mac users with large unified-memory machines
Perplexity's current local optimization work targets Apple silicon. The specific benchmark in its research uses an M5 Max with 128 GB unified memory, so users with smaller machines should not assume identical performance.
The direction still matters: Apple silicon's unified-memory architecture makes it increasingly practical to run useful local models alongside cloud systems.
Teams building AI agents
Agents are the strongest use case because they touch more data and run for longer periods than ordinary chat.
If an agent can browse files, read apps and retain memory, a routing policy becomes part of the security architecture. TowCue's AI agent best-practices guide remains relevant: local execution does not remove the need for permissions, stop conditions and review checkpoints.
Who should not overreact
Hybrid Compute is not automatically the best architecture for every task.
If your work is already public, low sensitivity and cloud-native, local routing may add complexity without meaningful benefit. If you need maximum model capability on every step, a smaller local model may also reduce quality on the portions kept on-device.
And "processed locally" should not be interpreted as "risk-free." Local agents can still access sensitive files, make mistakes or act with excessive permissions.
The right question is not whether local AI is safer in the abstract. It is whether the system gives you a clearer, enforceable boundary around the data that matters.
A low-risk way to test Hybrid Compute
If you have access to Perplexity Computer on Mac, test the privacy model before making it part of a real sensitive workflow.
- Start with a folder containing synthetic or non-sensitive documents.
- Ask Computer to combine local information with public web research.
- Observe which parts of the task remain local and when cloud escalation is requested.
- Test a document containing clearly fake personal data such as a sample phone number or email.
- Check whether the system keeps, redacts or asks permission for that information as expected.
- Only then move to a narrow real-world workflow with reversible consequences.
- Keep a fallback process for anything the routing layer handles incorrectly.
This is the same principle TowCue recommends in the AI tool selection guide: validate the workflow boundary, not just the model output.
TowCue take
The most important AI architecture shift may not be "local versus cloud." It may be local plus cloud, with routing as the product.
Cloud models remain better for many high-complexity tasks. Local models are increasingly good enough to handle private context, lightweight reasoning and device actions. The practical opportunity is to connect them without sending everything to the same place.
Perplexity's Hybrid Compute on Mac is an early example of that design becoming a user-facing product decision rather than an infrastructure experiment.
For privacy-sensitive knowledge work, that is worth watching closely. But the real test is not the phrase "on-device." It is whether users can understand and trust the routing boundary over long, messy, multi-step tasks.
Research sources
- Perplexity Blog — Introducing Hybrid Compute on Mac (September 1, 2026; official blog index)
- Perplexity Research — PII-TRACE: Detecting Personal Data Before It Leaves a Device (September 1, 2026)
- Perplexity Research — Optimizing On-Device Inference for Apple Silicon (September 1, 2026)