Perplexity has announced Hybrid Compute, a new feature that splits AI tasks between cloud-based frontier models and local LLMs running on a user's machine. The tool, integrated into the Perplexity Mac app, automatically detects sensitive content and prompts users to keep confidential data on their device. A new privacy classifier suggests which files should remain local, and users can review these decisions before processing begins.

Lawyers preparing case briefs and other professionals handling sensitive information are cited as key use cases. Local model options include Gemma 2B and two variants of Qwen's 35-billion parameter model, with one variant post-trained by Perplexity. The system requires no terminal access for installation and displays real-time CPU, GPU, and memory usage.

Users can queue tasks via iPhone and avoid token charges for local processing. Perplexity's Jon Staff noted that fully cloud-based models still produce superior outputs, but Hybrid Compute offers a balance between privacy, cost, and capability. Currently available only on Apple Silicon Macs running macOS 15, the feature requires at least 32GB of unified memory and is accessible to Pro, Max, and enterprise subscribers.