Perplexity Rolls Out Hybrid Compute on Mac, Keeping Sensitive Data On-Device

On September 1, Perplexity announced that its "hybrid compute" feature is now available to all users of the Mac app: tasks handled by the Computer agent can start in the cloud, but steps involving privacy-sensitive files (such as blood test reports, tax returns, or litigation materials) are automatically orchestrated to run on a local model on the Mac. Cloud and local work is ultimately merged on-device into a single output, and sensitive data never leaves the device. The feature is now live for Pro/Max/Enterprise users.

Confirmed

• Perplexity (@perplexityai) explained the division of labor: cloud models handle search, planning, and reasoning, while steps involving private files or sensitive information switch to a local model, keeping data on-device.
• CEO Arav Srinivas also open-sourced the PII (personally identifiable information) classifier that decides whether a task runs locally. It is based on Qwen3 with roughly 596 million parameters and is available on Hugging Face (perplexity-ai/pplx-pii-masking), used to detect private information before data leaves the device and to decide whether to process locally, upload to the cloud, or skip. A research blog post on the privacy-preserving local runtime was published alongside.
• According to @testingcatalog, local model options include a roughly 5.6GB Gemma (requires 16GB RAM), a roughly 17.4GB Qwen 2.5 32B (requires 32GB RAM), and Perplexity's in-house model (around 3B); a cloud orchestrator coordinates the work and dispatches suitable tasks to the local model.
• Srinivas explained the motivation: most AI agent apps don't distinguish between native and web runtimes, and the Mac offers an opportunity to shift token consumption to Apple Silicon—local tokens are free and privacy-preserving.

Why it matters

• Hybrid compute offers a new architecture balancing privacy and compute demands: it retains the reasoning power of frontier cloud models while keeping sensitive data local—a critical trust foundation for agent apps handling tax, medical, and similar documents.
• Open-sourcing the PII classifier lowers the barrier to replicating similar "cloud-device orchestration" approaches, and signals Perplexity's strategy of using Apple Silicon to cut cloud compute costs and position the Mac as a new platform for agents.

2026-08-31 ~ 2026-09-02 · 12 related posts

• Perplexity Mac Hybrid mode: Local models handle subtasks — testingcatalog · 2026-08-31
• Perplexity Hybrid details: Supports Qwen and Gemma locally — testingcatalog · 2026-08-31
• [source] Perplexity launches hybrid compute: cloud agents hand off sensitive steps to local Mac models — perplexity_ai · 2026-09-01
• Perplexity explains hybrid compute: sensitive data never leaves your Mac — perplexity_ai · 2026-09-01
• Perplexity open-sources PII classifier for hybrid compute privacy — perplexity_ai · 2026-09-01
• [source] Perplexity launches hybrid compute on Mac app, open-sources PII classifier — perplexity_ai · 2026-09-01
• Perplexity brings hybrid compute to Mac app: sensitive files handled by local models — AravSrinivas · 2026-09-01
• Perplexity CEO: local tokens on Apple Silicon are free and private — a big agent opportunity — AravSrinivas · 2026-09-01
• [source] Perplexity open-sources its PII classifier for hybrid compute, a ~596M-param Qwen3 model — AravSrinivas · 2026-09-01
• Perplexity publishes research blog on its privacy-protecting local runtime — AravSrinivas · 2026-09-01
• Perplexity Mac app adds hybrid compute for local file privacy — iamfakhrealam · 2026-09-01

1 near-duplicate retellings: inductionheads