Perplexity Hybrid Compute: Cloud Agents on Mac
Perplexity releases hybrid compute on Apple Silicon Mac for Pro, Max, and Enterprise subscribers, orchestrating cloud agents with a local model.
Stock photo for illustration only, not from the actual event
- Perplexity launches hybrid compute on Mac for Pro, Max, and Enterprise subscribers
- Tasks start in the cloud while sensitive data processing shifts down to the local Mac model
- An on-device classifier inspects files and automatically handles personal data security
- Enterprise admins can set organization-wide governance and view detailed audit logs
Perplexity has rolled out hybrid compute on Apple Silicon Macs for Pro, Max, and Enterprise subscribers running macOS 15 or later. The system requires at least 24GB of unified memory, with 32GB recommended. The local model installs with a single click directly from the Mac application, requiring no Ollama, no separate runtime, and no API key, while local execution consumes no cloud credits.
The orchestration direction defines the core architecture. Computer initiates every task in the cloud, where frontier models handle web searches, planning, and extensive reasoning. When a workflow intersects with private files or sensitive data, Computer routes that specific step down to the local model on the Mac without interrupting the task or dropping context, subsequently merging both halves into a cohesive result.
Stock photo for illustration only, not from the actual event
This approach reverses the local compute mode that Perplexity deployed on NVIDIA DGX Spark just a week earlier, which initiated tasks on user hardware and escalated to cloud models upon permission. Both paradigms share the exact same orchestrator but operate with opposite defaults.
Because Computer integrates with the iPhone, tasks can be triggered remotely while sensitive steps execute securely on a desktop Mac. Perplexity explicitly positions an always-on Mac mini as a dedicated local inference node tailored for this exact operational pattern.
"Before anything from a protected file reaches the cloud, an on-device classifier inspects it and the gate applies one of four outcomes."
Perplexity
Prior to transmitting any protected file data to the cloud, an on-device classifier inspects the content and enforces one of four distinct outcomes: keeping data local, masking sensitive spans, refusing the action, or requesting explicit user consent. Credentials, payment card numbers, and government identification receive the highest tier of security enforcement. Masked values are substituted with placeholders during transmission and restored upon the cloud response's return.
The integration of hybrid cloud-local computation marks a pivotal advancement in balancing high-performance AI reasoning with strict data privacy regulations. By delegating sensitive parsing tasks to on-device infrastructure while leveraging cloud scalability for general reasoning, platforms can mitigate enterprise compliance risks without sacrificing feature capabilities.
Source: MarkTechPost
Found something wrong in this article? Report an issue with this article
Comments
Leave a Comment