Perplexity is launching what is says is the first-ever hybrid inference orchestrator, debuting on Apple silicon. 

A Mac can now become the local half of one continuous Perplexity Computer task, working alongside frontier cloud models without the user switching modes, restarting the work, or managing the handoff.

Here’s how it works: the Mac’s computing starts in the cloud for its hardest reasoning, search, and planning, then calls on a local model running on Apple silicon when a step involves sensitive information. An on-device Privacy Gate checks prompts, tool outputs, memory, and logs before anything reaches the cloud. Users can even start a task from an iPhone and send its local steps to their Mac.

“For Mac users, the computer already on their desk is no longer just a window into cloud AI,” says Perplexity spokesperson Beejoli Shah. “Its Apple silicon becomes active compute inside the workflow, handling private steps locally while frontier models take on the rest. Apple mentioned Perplexity in its second quarter earnings call, and hybrid compute shows why the Mac is becoming a serious AI machine.”

Here are details from a Perplexity fact sheet:

° Truly hybrid inference requires a harness that can orchestrate models, files, and tools, as well as compute across the cloud and local device.

° Hybrid compute on computer begins in the cloud, using the most capable frontier models available. When a step involves sensitive data, Computer orchestrates down from the cloud to the device without restarting the task or losing context. This differs from local compute, which Perplexity launched on NVIDIA DGX Spark last week. Local compute starts on the user’s hardware and orchestrates up to frontier cloud

models only with permission.

° The cloud will handle frontier reasoning, web search, and long-horizon planning, while the local subagent handles private files, local data, and actions on the device.

° A Perplexity-trained Privacy Gate runs on the device and identifies sensitive information such as names, addresses, account numbers, and secrets. Computer flags that

information before anything is sent to the cloud, and the user decides where that portion of the task runs.

° Users can start a task from any Apple device, including an iPhone, and send local steps to their Mac without restarting. Computer completes the cloud and local work seamlessly

within one task, returning a single result. That opens up work most people would never hand to a cloud-only agent: privileged

legal documents, confidential client data, and patient records.

° At launch, users can choose from three local models: Gemma E4B, Qwen3.6 35B-A3B, a Perplexity post-trained version of Qwen3.6 35B, with more to follow. Setup takes one click inside the desktop app, with no terminal required.

° Users are not charged for any tasks that run locally. Charges only apply for frontier model usage.

° No new hardware is required to run hybrid compute. It works on any Apple silicon Mac running macOS 15 and higher.

° Due to the size of more advanced local models, it runs the best on devices with at least 32GB of unified memory.

° Enterprise admins can set one organization-wide sensitivity policy for the company and see a full record of what leaves each device.

Hybrid compute will be available through the Perplexity desktop app on Apple devices with Apple silicon.

I hope you’ll help support Apple World Today by becoming a patron. Almost all our income is from Patreon support and sponsored posts. Patreon pricing ranges from $2 to $10 a month. Thanks in advance for your support. 



Article provided with permission from AppleWorld.Today