Perplexity launched a fully local mode for Computer on Mac on October 20. You can now run the model and task execution on a supported Apple silicon Mac, rather than sending the job to cloud models. The first thing to check is your memory: Perplexity recommends 36 GB or more.
That wording matters for anyone comparing machines. It's a recommendation, not a published hard minimum. Don't buy a new Mac on the strength of the headline before checking support for the setup and jobs you actually need.
What runs on the Mac
Perplexity's announcement describes running "the AI model, harness, and task execution entirely on device". The harness is the software that directs the model's work and operates its tools. Local mode uses Qwen 3.6 or PPLX Pro, Perplexity's post-trained version of Qwen, and is available to Pro, Max and Enterprise subscribers.
The practical difference is that planning can stay on the device alongside file processing. Perplexity's earlier Hybrid Compute mode combines cloud reasoning with local handling of private information. It recommends less memory for that option, so check which mode you're selecting before treating the two as interchangeable.
According to Juli Clover's report at MacRumors, tasks involving current web information or cloud-connected apps can use cloud access with your authorization. Local mode therefore gives you a choice about where work runs. It can't make an online service operate offline.
Try it before upgrading your hardware
Start with copies of documents you already understand. Ask Computer to compare two versions and list the changes, then inspect the result against the files. Keep the first task small enough that checking it takes minutes. You're evaluating both the answer and how the assistant handles files.
Then try a job that changes something reversible, such as preparing a new document in a test folder. Check where it saved the result and whether it touched the originals. Running on your computer brings useful access; it also brings your everyday files within reach of mistakes.
Time the whole exercise, including your review. A local answer that needs extensive repair won't become useful just because the processor is yours. Equally, a modest model can be perfectly adequate for a narrow, repetitive task. Give it that task before asking it to reorganize your working life.
Privacy and ownership are separate choices
If you mainly want to avoid uploading particular files, local processing is an attractive option. Check permissions and any cloud handoff before using sensitive material. A native app can have online connections of its own, so the location of the model is only part of the journey a task takes.
If you want to operate independently of a subscription, this release doesn't promise that. You still use Perplexity's app and a qualifying plan. Our Muse Spark Open article looks at another route to local inference, with a very different hardware burden and license.
For an existing subscriber with a suitable Mac, a small local trial makes sense. For everyone else, wait for convincing results on the work you do. The upgrade worth paying for is the finished job, not the word Local in a menu.
Cover: AI-generated editorial illustration by TMRW using Perplexity branding. The hardware scene is illustrative, not an app screenshot.
