Perplexity launched Hybrid Compute, a Mac feature that splits one task across a frontier cloud model such as Opus 5 or GPT-5.6 Sol and a small model running on the user’s own machine, so sensitive material stays on the device. The company trained a privacy classifier that checks anything headed for upload and proposes what should stay local. Users get to review that list and choose which models handle which half before the work starts. Locally, the choices are Gemma E4B and two versions of Qwen’s 35-billion-parameter 3.6 model, one post-trained by Perplexity. Tokens a local model produces aren’t billed. Jon Staff, who runs Perplexity’s Mac products, said a fully cloud run will almost always produce a better artifact, and that the case for splitting comes down to privacy and cost instead.
