Perplexity Computer Shifts Only Sensitive Tasks to Local Mac Processing

Perplexity announced a new feature for its agent system on September 1 (local time): hybrid compute, added to Perplexity Computer. The core idea is that where the compute actually runs can change mid-task. Computer starts a job in the cloud, and once it reaches a stage involving personal files or sensitive data, it hands off to a local model running on the user's Mac. The feature went live in the Perplexity Mac app the same day it was announced.

To put it another way, Perplexity had already shipped a fully local "Portable Computer" on August 25, but that version only ran on separate hardware—NVIDIA's DGX Spark. This new hybrid compute is essentially a scaled-down version of that local processing capability, built to run on an ordinary Mac without any extra hardware.

Why the cloud alone wasn't enough

Perplexity Computer is an agent that autonomously handles multi-step tasks, from web research to file organization. The trouble is that along the way, some of those steps touch personal documents or internal company data that users would rather not send to the cloud. Perplexity has been working through this problem in stages over the past month. On August 25, it launched a fully local "Portable Computer" that ran the entire orchestrator, sub-agents, and agent harness on local hardware—NVIDIA's DGX Spark. The very next day, August 26, it introduced "Brain," a self-improving memory system that organizes sessions and files into a knowledge wiki. Hybrid compute can be seen as bringing that same local-processing capability to an ordinary Mac, without requiring dedicated hardware like DGX Spark.

PPLX Qwen 3.8 27B and the privacy gate

According to TestingCatalog, the local model powering hybrid mode is PPLX Qwen 3.8 27B, a model Perplexity post-trained in-house. The "27B" refers to 27 billion parameters—far smaller than the large frontier models running in the cloud, small enough to run on a regular Mac. Alongside it comes a "privacy gate" feature that filters out personally identifiable information (PII) before it's sent to the cloud. Rather than requiring users to decide each time when a task should move local, the gate appears to detect sensitive data and trigger the switch automatically.

Perplexity Computer now comes in three flavors

Over the past month, Perplexity has expanded the same underlying agent into three variants, depending on where it runs.

Version	Runs on	Model used	Availability
Computer (original)	Cloud	Frontier cloud models	All users
Portable Computer	NVIDIA DGX Spark (local hardware)	PPLX 27B / Qwen 3.8 27B	Perplexity Pro/Max subscribers
Hybrid Compute (new)	Cloud → switches to local Mac	PPLX Qwen 3.8 27B	Perplexity Mac app

All three rely on the same family of local models, but access differs depending on which hardware they're paired with. Where the DGX Spark version cuts cloud dependence entirely, hybrid compute is a middle ground: it leans on cloud power for most of the work and narrows down to local processing only when needed.

Cloud side gets Claude Fable 5.1

On the same day, Perplexity issued a separate announcement updating the cloud side of Perplexity Computer as well. Anthropic's Claude Fable 5.1, released that same day, is now available directly in Computer for Pro and Max subscribers. In Perplexity's own August WANDR benchmark, Fable 5.1 topped the leaderboard with a score of 0.601 at a cost of $12.76 per task. Compared with the previous generation, Fable 5 (0.496 points, $20.30 per task), that's a 21% higher score at 37% lower cost.

The accompanying chart also included scores for other models tested on Computer.

Model	WANDR Score	Cost per Task
Fable 5.1	0.601	$12.76
Opus 5	0.537	$11.60
Grok 4.6	0.496	$7.58
GPT-5.6 Sol	0.426	$4.99
GPT-5.6 Terra	0.399	$1.98
DeepSeek V4 Pro 0813	0.359	$0.75
Fable 5	0.496	$20.30
Sonnet 5	0.309	$5.75

So while adding a local 27B model to handle sensitive steps, Perplexity also widened its cloud-side model lineup to include the newest top-tier model released that same day. That's why the two announcements landed together.

How to use it

Hybrid compute attaches automatically whenever you hand a task to Computer in the Perplexity Mac app. The task starts in the cloud, and once it reaches a stage involving personal files or sensitive data, it switches over to the local PPLX Qwen 3.8 27B model without any manual action needed. At that point, the privacy gate detects PII and keeps it from being sent to the cloud. For example, if you ask Computer to organize contracts or HR files sitting in a local folder, the early steps that require web search happen in the cloud, while the part that actually reads the documents runs entirely on your Mac. The announcement doesn't specify whether the terms differ across subscription tiers.

Official video

Editor's take

Looking at how Perplexity split Computer into three tracks—cloud, fully local, and hybrid—within the space of a month, it's clear the problem the company is trying to solve isn't performance. It's trust. For an agent to take over reading and organizing files, users have to hand those files over, and Perplexity clearly knows plenty of people don't want their tax documents or HR records going anywhere near the cloud. The DGX Spark version answered that concern with hardware; hybrid compute answers the same concern with a software-level switch instead.

Anyone who's put a local model this size to real work tends to reach the same conclusion: a 27B-class model can't match the reasoning depth of a cloud frontier model, but it's more than capable of reading and summarizing documents that are already well-structured. That's exactly why the hybrid approach makes sense—let the cloud handle the hard judgment calls, and let the local model handle the simpler work of actually looking at sensitive material.

For companies weighing whether to attach an agent to work involving contracts or HR data—material they'd rather not send outside—this is worth watching. That said, how tightly Perplexity has tuned the cutoff point between cloud and local processing is something you'll only really get a feel for after using it for a few days. The broader the privacy gate's definition of PII, the safer it is, but that also means tasks will fall back to local processing more often.

The open questions for the coming weeks are whether this local model family expands beyond Mac, and whether Portable Computer's DGX Spark-only architecture eventually merges with hybrid compute. Given that Perplexity keeps experimenting with running the same 27B model across different hardware, the company's endgame looks like an agent that can handle most tasks without needing the cloud at all.