Technology
Perplexity introduces Hybrid Compute to split AI tasks between cloud and local models
The new Mac feature aims to protect sensitive user data and reduce inference costs by utilizing local hardware.
The short version
- Perplexity has announced Hybrid Compute for its Mac application, allowing users to divide tasks between cloud-based frontier models and local large language models.
- A newly trained privacy classifier automatically scans uploads to suggest which files and sensitive details should remain secure on the user's local machine.
- The feature is currently restricted to Pro, Max, and enterprise subscribers using Apple Silicon Macs with macOS 15.
Key facts
- Perplexity's Hybrid Compute splits tasks between cloud-based frontier models, such as Opus 5 or GPT-5.6 Sol, and local models running on the user's computer.[Engadget]
- The tool is designed to protect sensitive personal data by processing it locally while also reducing overall inference costs.[Engadget]
- Local model options currently include Gemma E4B and two variants of Qwen's 35-billion parameter 3.6 model, with installation handled directly through the Perplexity app.[Engadget]
- The feature is limited to Apple Silicon Macs running macOS 15, with Perplexity recommending at least 32GB of unified memory.[Engadget]
What remains uncertain
- Jon Staff, Perplexity's lead for Mac products, acknowledged that a fully cloud-based system will almost always produce better raw output quality, meaning the actual performance compromise for users remains to be seen.[Engadget]