Perplexity brings its full Computer agent to local Apple Silicon Macs

Perplexity now lets Pro, Max and Enterprise subscribers run its full Computer agent on an Apple Silicon Mac with at least 36GB of unified memory, removing the cloud leg of its Hybrid Compute offering released at the start of September.

The new mode arrives one month after Perplexity shipped Hybrid Compute, which split workloads between local hardware and remote frontier models (large, expensive cloud AI systems). Hybrid Compute still assumed a cloud round trip. The new option drops it.

Perplexity framed the move around two pressures: privacy concerns and credit spend among heavy users. "As more demanding workflows become practical to run locally, the machines people already own will take on a larger share of their AI work," the company wrote. "This gives users more control over sensitive data and credit spend."

What did Perplexity actually move onto the Mac?

Until today, running Perplexity Computer locally meant borrowing the device as a stepping stone. The harness — the routing layer that decides which model handles each step — sat in the cloud. Tool calls, code execution and final synthesis all required a round trip to Perplexity's servers.

Now both the harness and orchestrator (the workflow coordinator that sequences multi-step tasks) execute on the user's Mac. The company says this lets the agent touch files in Notes, Mail and other local apps without that data leaving the device.

A practical example surfaced in the announcement. With access to Notes and Mail, a user can ask Computer to generate a priority to-do list drawn from inbox items. No email content travels to Perplexity's servers to answer the question. Cross-app synthesis at that scale — small enough to run on-device, useful enough to matter — is the working definition of a useful local agent.

What does the 36GB floor leave out?

Three caveats follow the promise.

First, the gate. The 36GB unified-memory floor excludes most older and entry-level Apple Silicon machines. Apple's standard laptop and desktop configurations ship with much less. The minimum spec targets Apple's higher-end MacBook Pro and Mac Studio configurations — machines that start in the multi-thousand-dollar range.

Second, the model menu. Users can pick between Qwen 3.6 or PPLX Pro. The latter is a Perplexity post-trained variant of Qwen 3.6, fine-tuned on Computer commands. Perplexity has not disclosed parameter counts (model size, measured in billions of weights), quantization (a compression technique that shrinks models at some quality cost) or context window sizes (how much text the model can process at once). All three numbers decide whether the model actually handles a real workload.

Third, the sandbox. Code and tool execution continue to run in an isolated sandbox (a restricted environment that limits what code can access). Perplexity confirmed the framing: "These same security protocols persist in the cloud version of Perplexity Computer." Sandbox isolation is a baseline, not a zero-risk claim. The release notes do not specify whether the local sandbox lives on the Mac or in the cloud for fully local sessions.

Does the privacy claim hold up?

The privacy pitch demands more than the announcement offers. A local harness does not guarantee local telemetry (diagnostic usage data sent back to the vendor), local crash logs or local caching of intermediate outputs. Perplexity did not publish a data-flow diagram, audit summary or penetration-test reference (an independent security audit that probes for vulnerabilities).

For enterprise customers subject to HIPAA (US health-data law), GDPR (EU data-protection regulation) or SEC scrutiny, marketing claims about "staying on device" do not replace contractual terms. A privacy-focused subscriber should treat local execution as a default-deny configuration and verify with packet captures.

The 36GB requirement also hints at what runs beneath. Qwen 3.6 in full precision would far exceed 36GB in any reasonable form. Perplexity almost certainly ships a quantized variant (a compressed model with lower numerical precision, trading accuracy for footprint). The company did not publish the quantization level used. That is the single most important technical detail for any user weighing model quality.

What role does Hybrid Compute play now?

Local-only is not the only path. Subscribers can authorize the agent to call on more than 15 hosted frontier models when a task exceeds local capacity. The routing decision still sits with the local harness, which decides per step whether to call the device or the cloud.

This hybrid posture matters. Some tasks — long document synthesis, large code refactors, multi-file analysis — still punish local inference (the on-device generation step that produces model outputs). The fallback to frontier models keeps Computer usable on workloads no 36GB machine can crack today.

That position mirrors a broader industry tilt. Apple Intelligence, Microsoft Copilot+ and Qualcomm's Snapdragon X platform have all pitched on-device AI as a privacy and cost advantage. None has shipped a fully local agent at this scale for general consumers.

Who can use it today?

Three subscription tiers unlock the feature on launch day: Perplexity Pro, Max and Enterprise. After installing the latest Perplexity Mac app, subscribers follow in-app prompts to download and install a chosen model. Free-tier users wait.

What is Perplexity betting on upstream?

The company's local offering leans on Qwen, an open-weight model from Alibaba. Perplexity built PPLX Pro on top of that base, but the underlying weights remain upstream. If Alibaba tightens licensing or ships a competing consumer agent with deeper Mac integration, Perplexity loses leverage overnight.

So what: Local-agent capability is moving from demonstration to shipping feature, and Perplexity just handed Apple Silicon owners a meaningful privacy tool — provided they own the most expensive half of the lineup. The next quarter will show whether rival labs match the 36GB threshold on cheaper hardware, or whether local AI simply becomes a new premium tier only top-shelf Macs can afford.