Perplexity AI Updates: September 2, 2026

1. Perplexity Ships Hybrid Compute on Mac

Perplexity. The Mac app can now add a compact local model post-trained for Perplexity Computer and split each task between the device and the cloud. A local privacy gate classifies names, addresses, account numbers, and secrets before anything leaves the machine, then keeps that content on device, redacts it, or asks for approval. Supported local models are Gemma 4 on 16 GB machines and Qwen3.6 or Perplexity’s own model on 32 GB machines. Source

2. PII-TRACE Benchmarks PII Detection Across Long Conversations

Perplexity. PII-TRACE, short for Tracing Recurring PII Across Conversational Exchanges, is a 13-language benchmark for evaluating whether PII detectors stay consistent as identifiers recur across a long conversation rather than appearing once in isolation. Perplexity paired it with PII-Tracer, a 0.6B detector small enough to run locally, which is the model behind the hybrid compute privacy gate. Publishing the benchmark alongside the detector makes the privacy claim in the Mac release checkable. Source

3. Perplexity Details Its Custom Apple Silicon Inference Engine

Perplexity. The engineering post covers a custom local inference engine for Apple Silicon that improves both prefill and decode throughput over off-the-shelf runtimes. Throughput is the binding constraint for the hybrid compute feature, since the local model has to keep up with interactive use before users will accept keeping work on device. Source