For much of the past year, the practical limit on AI agents that touch personal files has been the boundary between the device and the cloud.
Tasks that needed recent information, browser access, or stronger reasoning stayed online. Anything involving private documents, messages, or local applications raised the usual questions about what left the machine and when.
Perplexity's Computer is able to work across local files and Mac apps through its Personal Computer mode.
This time, Perplexity has added an option in which both the model and the agent that directs the task run entirely on the Mac.
In this local mode, planning, reasoning, and file processing stay on the device.
The app offers two local model choices that users download through the Perplexity Mac app: Qwen 3.6 and PPLX Pro, a version of Qwen that Perplexity has post-trained for its Computer workflows.
With permission, the agent can read and act on local files and operate across supported Mac applications such as Notes, Messages, Mail, Numbers, and Keynote. One example given is reviewing recent messages and emails, identifying deadlines, and producing an action list in Notes.
Code and tool execution occur inside isolated sandboxes with controlled access to files and apps, the same security approach used in the cloud version of Computer.
Local tasks do not consume Computer credits.
Users can still authorize cloud access when a task requires current information, browser use, connected services, or the stronger reasoning available from one of more than fifteen frontier models.
Hybrid Compute remains available in the same app.
In that mode, cloud models handle research and longer reasoning while a local model on the Mac processes private files; an on-device classifier flags potentially sensitive information so the user can choose whether to mask it, keep the step local, or allow it to proceed.
The fully local option simply starts and finishes the entire task on the Mac instead.
The feature is available now to Perplexity Pro, Max, and Enterprise subscribers on supported Apple silicon Macs running macOS 15 or later.
Perplexity recommends 36 GB or more of unified memory for the local mode.
Hybrid Compute works on machines with as little as 24 GB. Purely local work can proceed without an internet connection for the model inference itself, though any step that sends email, checks a live service, or reaches a connected app still needs network access and explicit approval.
As for why Mac-specific hardware and not others, is because several hardware and software characteristics are harder to replicate elsewhere at the same level of integration.
In this case, Apple silicon's unified memory lets larger models remain resident without the data movement and capacity constraints typical of discrete GPU setups, which is why a model the size of the Qwen variants can run usefully on the device once memory is sufficient.
Perplexity has also built a custom inference engine, Lily, tuned specifically for Apple silicon's Neural Accelerators, Metal stack, and memory architecture.
Tthe company has published results showing higher prefill and decode throughput on this path than on general frameworks such as MLX for the relevant Qwen configuration.
On the software side, the Mac already provides deep, permissioned access to native apps and a continuity layer that lets tasks start from an iPhone while private file work stays on the Mac.
A Mac mini can remain on as a dedicated local node without tying up a primary laptop. Other platforms have their own local agent options, like Perplexity's Portable Computer that can run the model and agent on supported Nvidia and AMD hardware under Windows or Linux.
However, the combination of unified memory, the custom Apple silicon runtime, and direct native app control is what the current fully local Computer mode is built around.
