Perplexity Computer now runs locally on Mac with Qwen 3.6

What's going on

AI company Perplexity has extended its macOS agent system Perplexity Computer. The app controls a Mac, remotely if you like. Until now, this required a cloud language model, and therefore a constant internet connection. The app can now also work with local LLMs (large language models) that run directly on your machine, heise reports.

Which models run locally

Two models are planned for launch:

• Qwen 3.6, a language model from the Qwen family.
• PPLX Pro, a Qwen model that Perplexity developed further through post-training. Post-training means fine-tuning an already pretrained model for specific tasks.

According to heise, it's unclear whether more local models will follow, and how well the models are optimized for the task.

Requirements and pricing

The Mac must be an Apple Silicon device. Perplexity recommends at least 36 GB of RAM and names newer machines such as the current Mac mini lineup with M6 and M5 Pro. But heise says the M6 compact Mac currently doesn't come with more than 32 GB. You'd need an M5 Pro model with 48 GB, which costs at least €2,659.

It isn't free: only customers with a Pro, Max or Enterprise subscription get access. You download the app from perplexity.ai. It isn't available in the App Store.

What should be possible locally

Perplexity cites "complex multi-step work" that can be done entirely locally. Examples include analyzing your private financial portfolio and working with files, notes, the Messages app, spreadsheets and Apple Mail. If the app runs fully locally, no data should leave your computer. That's a privacy advantage, especially for sensitive content like finances and email.

Hybrid Compute

If local performance isn't enough, you can bring in the cloud. With "Hybrid Compute," a local orchestrator (a control layer that distributes tasks) activates one of more than 15 frontier models in the cloud. Perplexity says: "How the user spends credits is up to them." For remote control of the Mac, Perplexity says code and tool execution run in isolated sandbox environments. That's a vendor claim that heise hasn't verified.

Analysis

Local agents are appealing because your private data stays on the device and there's no cloud dependency. The hurdle is hardware: 36 GB of RAM and Apple Silicon rule out many users, and you also need a subscription. It remains to be seen how smoothly switching between the local model and the cloud works, how reliably the local models handle computer tasks, and whether the sandbox promises hold up. According to heise, tests still have to show that. The app itself is proprietary, and the Qwen 3.6 models are just an ingredient here.