The Mac app now runs Computer tasks without an internet connection. The model and agent handle planning, reasoning, and file work on the device, and files go to the cloud only after user approval.

The update ships with Qwen 3.6 and PPLX Pro as downloadable local options. It is open to Pro, Max, and Enterprise subscribers.

Hybrid Compute remains in the same app, so the local path is an added route rather than a replacement. The 36GB unified memory requirement puts the useful path on Macs with enough memory to hold the model and agent.

Local processing can keep files from automatic cloud transmission, but the pages do not state model weights, price, latency, token cost, or benchmark scores. The claim is a deployment path, not a performance proof.

The 36GB memory floor makes this a capacity decision, not a node decision. The filing establishes a local deployment option, not a speed, quality, or cost result.