Perplexity just shipped something I've wanted from an AI company for a while now: Hybrid Compute, a feature that splits work between cloud models and smaller models running on your own machine, depending on how sensitive the task is.

The core idea is simple. Work gets divided between cloud and local models based on the task. In principle, a throwaway question about dinner plans could go cloud-side where the big models live, while anything involving your personal documents or other details you'd rather not ship to a datacenter stays local, handled by a compact model running on your hardware. The specifics of how Perplexity classifies each request haven't been laid out in detail yet, so treat my examples as illustration rather than documentation.

This maps almost perfectly onto something I deal with at work constantly. Every distributed system I've ever touched has some version of this tradeoff: do you send the work to the big central service, or handle it at the edge where the latency is zero but the compute is weaker? Hybrid Compute is basically an edge-computing pattern applied to your prompts. And to me the interesting engineering isn't even the models themselves, it's the router deciding which lane each request goes down. Too aggressive and you get bad answers on questions that deserved the big model. Too conservative and the privacy pitch falls apart.

The interesting question in 2026, at least in my view, is what a model can do without a datacenter holding its hand. Local inference used to be a novelty demo, a party trick you ran with llama.cpp to show your CPU melting. Now smaller models are genuinely good at a lot of routine tasks, good enough that running them locally for sensitive stuff isn't a compromise anymore.

Most of what people ask AI assistants probably doesn't need a frontier model anyway. You don't burn a trillion-parameter model to summarize an email or pull a date out of a receipt. The cloud should be reserved for the hard stuff.

The catch, and there's always a catch, is transparency. When a request gets routed local versus cloud, does the user see it? Can you override the router and force everything local when you're having a paranoid day? A hybrid system only builds trust if the split is visible instead of a black box, and that's the part I'll be watching as this rolls out.

Still, this is the right direction. Privacy shouldn't require giving up model quality entirely, and quality shouldn't require handing over everything. Splitting the difference is harder than picking one lane, but it's what actually matches how people use these tools. Would love to see Apple and Google copy this pattern wholesale.