Split tasks between cloud and local models for data privacy

A single agent task is intelligently divided. Cloud models handle research and heavy reasoning, while a compact local model on your Mac processes any step that involves private data. This ensures sensitive information is never transmitted.

Business outcome: Enables use of AI on confidential data, reducing legal and compliance risks associated with cloud processing.

Local detection of sensitive information with user control

A local classifier detects sensitive information like names, emails, and account numbers. It prompts the user to keep that slice of the job on-device, mask the details, or send it to the cloud anyway, giving the user control.

Business outcome: Reduces the chance of accidental data leakage by providing a clear, auditable choice at the moment sensitive data is encountered.

Transparent and auditable security with an open-source PII model

Perplexity open-sourced the on-device PII classifier, developed with its Secure Intelligence Institute. It classifies nine PII categories, tracks their boundaries in text, and rates the sensitivity of the whole conversation, offering transparency.

Business outcome: Offers a higher level of trust and allows security teams to audit the very mechanism responsible for data protection.

Choose the on-device model that fits your Mac's hardware

At launch, users can select from local models including Gemma E4B, Qwen3.6 35B-A3B, and PPLX Qwen 3.8 27B. The choice depends on what their Mac can handle, offering flexibility between performance and resource usage.

Business outcome: Allows teams to tailor the balance between on-device capability and hardware constraints, optimizing for their specific Mac fleet.

Reduce cloud serving costs by using local hardware

Tokens processed locally on the user's own machine are not billed. This shifts a portion of the inference workload onto user hardware, which independent commentary notes reduces Perplexity's serving costs.

Business outcome: Provides a direct economic benefit for high-volume users by lowering the number of billable tokens consumed on cloud models.

Available on paid tiers for professional and enterprise use

Hybrid Compute is available to Pro, Max, and enterprise subscribers. This ensures the feature is accessible to individual professionals and can be rolled out across an organization with a managed subscription.

Business outcome: Provides a clear pathway for businesses to standardize on a secure AI agent solution with predictable per-seat costs.

• Data Privacy on Sensitive DocumentsIts core strength is allowing AI agents to work with confidential files without the data ever leaving the user's Mac, a major advantage for regulated industries.
• User Control and TransparencyThe PII classifier prompts users for action on sensitive data, providing a clear audit trail and control over what is processed where.
• Open-Source Security MechanismOpen-sourcing the PII classifier demonstrates a commitment to transparency, allowing independent security review and building user trust.
• Economic EfficiencyBy processing tokens locally, it reduces the number of billable cloud tokens, offering a cost advantage for users with high-volume, data-sensitive tasks.

• Mac-Only at LaunchThis feature is exclusively for the Mac app with no announced support for Windows or Linux, limiting its utility in mixed-OS organizations.
• High Memory RequirementsA hard memory floor excludes most base-model Macs. There is a discrepancy in reporting: Perplexity states 24GB unified memory, while several outlets report 32GB as required or recommended for larger local models.
• Reduced Local Model CapabilityThe local models are materially smaller and less capable than frontier cloud models, so the quality of output on locally-handled steps will be lower than cloud-only processing.
• Professional RealityThe privacy benefit depends entirely on trusting the local classifier to correctly identify all forms of sensitive data. If it fails, information could be sent to the cloud unintentionally, and the paid-tier requirement adds a cost barrier.

Legal Document Review

A law firm can use the Computer agent to analyze a contract. The local model processes clauses containing privileged information, while the cloud model provides general legal research, ensuring client confidentiality.

Financial Data Analysis

An analyst can instruct the agent to summarize a portfolio's performance. Account numbers and personal client names are handled on-device, while market trends and economic data are researched in the cloud.

Healthcare Record Summarization

A medical administrator can use the agent to draft a summary of a patient's file. The local model processes the patient's name and contact details, while the cloud model assists with formatting and general medical knowledge.

Board Meeting Preparation

An executive assistant can have the agent prepare a briefing document from confidential strategy files. The local model handles the sensitive corporate data, and the cloud model helps structure the final report.

Ensure you have an Apple silicon Mac with macOS 15 or later and at least 24GB of unified memory (32GB is recommended for larger local models).

Subscribe to a paid Perplexity plan (Pro, Max, or Enterprise) to unlock the Hybrid Compute feature in the Mac app.

In the Mac app, navigate to the settings for the Computer agent and select your preferred local model from the supported list (e.g., Gemma E4B, Qwen3.6 35B-A3B, or PPLX Qwen 3.8 27B).

Start a task with the Computer agent; when the local PII classifier detects sensitive information, choose to keep it on-device, mask it, or send it to the cloud.