Perplexity Computer Now Runs Fully Local on Mac: What Really Stays On-Device

Perplexity has added a local-first mode to Computer, its agent, in the Mac app. The company says the model and the agent that directs it both run on the Mac, so planning, reasoning and file processing can happen on-device, including without an internet connection. It was announced on October 20, 2026 and is available now to Pro, Max and Enterprise subscribers.

This is a real step beyond September's Hybrid Compute, which split work between cloud models and a local model. But "fully local" comes with qualifiers, a memory requirement and a cost model that readers should understand before downloading anything. Sources: Perplexity's announcement on X as embedded by 9to5Mac, and MacRumors.

TL;DR: claims and fine print

What changed since Hybrid Compute?

In September, Perplexity shipped Hybrid Compute: frontier cloud models for research and reasoning, plus a local model that handles private files, with a small classifier that swaps personal data for placeholders before anything goes out. We covered that in Perplexity Mac hybrid compute and local PII.

Local-first goes further. 9to5Mac notes it "runs both the AI model and the agent directing the task on the Mac," with files reaching the cloud only when the user approves. Hybrid Compute remains available in the same app. Perplexity's example task: review recent messages and emails, identify deadlines, and write an action list in Notes.

Fully local, with exceptions

The phrase "fully local" in the headline and in the launch post should be read with the sentences that follow it in the coverage.

• If a task needs current information, browser use or cloud-connected apps, you can authorize cloud access and decide when to spend credits (MacRumors).
• You can call on more than 15 frontier models for tougher problems.
• Any task that touches the internet, naturally, needs a network connection.

So the accurate summary is: the default path keeps your files and reasoning on the Mac, and anything that leaves the machine is a deliberate, approved step. That is a meaningful privacy improvement, but it is a permissioned design, not an air gap. For comparison with cloud-first agents, see our coverage of Perplexity Computer projects and multiplayer and of how AI agent privacy promises compare.

Why 36GB?

A secrets key vault, illustrating Perplexity Computer keeping files and keys on the Mac in local mode

Running a capable agent model locally means holding its weights, a long context cache and the app in unified memory at once. Perplexity recommends 36GB or more, and notes Hybrid Compute needs only 24GB because the heavy model stays in the cloud. This tracks what we have seen in other local stacks: small quantized models fit anywhere, but agent workloads with tool calls and long contexts want headroom. Our guides to Perplexity's Lily inference engine for Apple silicon, llama.cpp Metal speculative decoding on M3 Ultra and the DwarfStar DS4 local engine show how much runtime choice affects speed on the same Mac.

Perplexity also highlighted the new Mac mini machines in its launch material, a sign that it sees small desktops as home agent servers. For its related hardware plans, see the Perplexity Portable Computer posts.

How well does it work?

There are no published benchmarks for local Computer. The only performance reports we found were readers. A MacRumors commenter with a 64GB M5 Pro Mac said they tried the 27B model to fill a presentation template for six events; they described a run of about 50 minutes, heavy battery drain and full GPU load, and said it finished without producing anything, where a cloud coding agent would have taken minutes. Another reader pointed out that the underlying open model can be run for free with other tools, questioning the subscription.

Both comments are anecdotes from anonymous commenters, not tests, and we have not reproduced either. They do show the questions to ask:

1. Latency and battery: local agents are slower and draw more power than cloud calls.
2. Reliability on long tasks: an agent that loops locally can fail without a clear error.
3. Value: if the local model is open weights, what is the subscription buying? Presumably the agent harness, native app integrations, sandboxing, the post-trained PPLX Pro model and the cloud fallback.

Privacy and security considerations

• Permissions: the agent can work with Notes, Messages, Mail, Keynote and Numbers if you grant access. Messages and Mail hold sensitive data; review the permission prompts rather than clicking through.
• Sandboxing: MacRumors reports code and tool execution run in isolated sandbox environments. How strong the isolation is has not been independently assessed.
• Prompt injection: a local agent reading email can still be steered by a malicious message. Local processing reduces data exposure, not instruction-following risk. If you run agents against inboxes, an action-gating layer such as AgentBeam, from the explainx.ai team, is built for exactly that problem.
• Cloud escape hatch: every time you authorize the cloud for a task, the privacy guarantee for that task changes.

How to try it and test it fairly

MacRumors describes the setup in two steps: download Perplexity for Mac, then choose the fully local option. 9to5Mac adds that the local models, Qwen 3.6 and PPLX Pro, are downloaded and set up through the app. Because there are no vendor benchmarks, a short test plan will tell you more than the launch post.

1. Pick three representative tasks. One that reads local files (summarize a folder of PDFs), one that spans apps (the Messages, Mail and Notes deadline example Perplexity gives), and one that needs the web, so you can see exactly when the app asks to leave the device.
2. Time each run and watch Activity Monitor for memory pressure, GPU load and thermals. Note battery percentage per minute if you are on a laptop.
3. Run the same tasks in Hybrid Compute and in the cloud mode. Compare answer quality, time and credits spent.
4. Disconnect from the network and repeat the first two tasks to confirm they truly work offline.
5. Check what the permission dialogs say each time the agent wants Mail, Messages or file access, and keep a record of any cloud escalation.

If the local model stalls or loops on a long task, as one reader reported, the practical fallback is the app's option to call a frontier model for the hard step rather than letting the local one run for an hour.

A short timeline

• Earlier in 2026: Perplexity ships Personal Computer for subscribers on Mac, covered by 9to5Mac alongside an "all-new native Mac experience."
• September 2026: Hybrid Compute arrives, splitting tasks between cloud models and a local model for private files (our September coverage).
• October 20, 2026: the local-first mode launches with Qwen 3.6 and PPLX Pro, and a 36GB recommendation.

The pattern is a steady move of more of the agent loop onto the user's machine: first the sensitive data step, now planning and reasoning as well. It echoes the direction of Perplexity's portable local-agent hardware and of Apple-silicon inference work such as its Lily engine.

What this means for the local-AI argument

Local agents are often sold on three arguments: privacy, cost and availability. Perplexity's launch speaks directly to the first and third (data stays on the Mac, and local tasks work offline) while the second is less clear. A subscription is still required, and local runs trade cloud fees for electricity, wear and time. For people who already own a 36GB or larger Mac, marginal cost per task may be low. For everyone else, the hardware requirement is the real price of admission.

The other open question is quality. Open models of the Qwen 3.6 class are good, and Perplexity says it post-trained one for Computer, but agent tasks are long chains where a small reliability gap compounds. A model that is 90% reliable per step is far less reliable across twenty steps. That is why a cloud fallback is not a footnote here, but central to whether the feature feels usable.

Who should try it

• Good fit: privacy-sensitive users with 36GB or more of memory, who want an agent working over local files and Mac apps, and are comfortable with slower runs.
• Maybe: 24GB Mac owners should start with Hybrid Compute.
• Poor fit: anyone who mostly needs live web research, where the cloud path will dominate.

What to watch

• Independent speed and accuracy tests against cloud Computer on common tasks.
• Whether Perplexity publishes details of the PPLX Pro model.
• How credits are metered when a local task escalates to the cloud.
• Windows or Linux equivalents: Perplexity's portable computer posts suggest that direction.

Related reading

• Perplexity Mac hybrid compute and local PII
• Perplexity Lily on Apple silicon
• Perplexity Portable Computer with DGX Spark
• Perplexity Computer projects
• AI agent privacy promises
• Build a personal AI system locally
• Official: 9to5Mac report and MacRumors report

Details are accurate as of October 22, 2026 and come from press coverage of Perplexity's announcement; the product is changing.