Perplexity introduces Hybrid Compute to blend cloud AI with on-device models

Perplexity announced Hybrid Compute, a capability that splits AI workloads between a frontier cloud model—examples being Opus 5 or GPT-5.6 Sol—and a locally hosted large language model on the user's computer. The design aims to keep confidential information on-device, addressing privacy concerns while also reducing the expense of cloud inference. Users can choose from local models such as Gemma E4B and two flavors of Qwen-3.6, one of which Perplexity has further fine-tuned.

Installation is handled through the Mac application without requiring terminal commands, and the interface displays CPU, GPU, memory usage and token consumption. Hybrid Compute is available to Pro, Max, and enterprise subscribers on Apple Silicon Macs with macOS 15 and a minimum of 32 GB unified memory. Jon Staff, who oversees Perplexity's Mac products, noted that while cloud-only outputs are generally higher-quality, many users prioritize data security and cost savings.