Introducing hybrid compute in Perplexity Computer. Computer can start a task in the cloud, then move to a local model that runs on your Mac. Use it for steps involving private files or sensitive data. Available today in the Perplexity Mac app.

Hybrid compute uses cloud models for search, planning, and reasoning. It orchestrates to a local model on your Mac to handle private files or sensitive information. Computer combines their work into a single output without ever taking sensitive data up to the server.

A privacy gate on your Mac detects PII before it leaves your device. Choose to run that step locally, send the file to the cloud, or skip entirely. We are open-sourcing the classifier today. Read the research: Huggingface:

Hybrid compute is available today for Perplexity Pro, Max, and Enterprise subscribers in the Perplexity Mac app with macOS 15+. Get started:

finally a reason for my macbook fan to sound like a jet engine while searching where to get lunch 💀

IVE BEEN WAITING A FUCKING WEEK FOR A RESPONSE

Who did the animations and design the website needs a raise!! brilliant job

STOP CHEATING WITH USERS! you set Computer as default and i need every time swithc to Search! you boost your revenue and steal money!!!!

Moving sensitive steps to a local model is the part I like most here. Hybrid compute could make AI assistants much easier to trust for real work.

Hybrid compute isn't just a trend; it's a massive shift in how we work. The real win isn't picking between the cloud or your own device. It's building systems that move between them. But they've got to stay private and verifiable.

30+ hours into a sandbox infrastructure failure on @perplexity_ai Computer that has trapped 13 production scheduled tasks I can't delete or modify. Live support unresponsive. Session 019ef01f. Paying customer with time-sensitive subscriber deliveries this weekend. Anyone home?

Would be a great Hermes feature 👀 @Teknium @NousResearch

Open-sourcing the 0.6B gate is the right move. The failure mode I care about is consistent identifiers across a long Computer thread — same client, 8 turns later, slightly paraphrased.

You can start a task but not finish it with the monthly 10,000 credits. Only usable for research to feed agents

It's genuinely funny that you guys are making 10th announcement of this thing

my fan starts screaming the second i fire up llama3 8b, does this actually throttle for thermal loads

Idk why people have forgotten about you. You are good.

using local models for sensitive data makes this much more practical

Great idea! Thanks!

Is this fully deployed? Have all the requirements and not seeing hybrid or the ability to download models on an M4 Pro. Support hasn't been responsive.

Hybrid compute lets steps shift from cloud to local models on your Mac.

if this is the daily driver, where does a human still have to catch it?

Hybrid compute is a smart move. Cloud-to-local switching for sensitive data could make AI agents much more practical 🔥

Really like the trend that systems offer more control on your data and their actions. When there is more clarity what should and should not happen, people will have more trust.

Give every run a simple audit trail: which step stayed local, what went to the cloud, and why. That's what would make the privacy promise easy to inspect.

The privacy gate is the product. Cloud for reasoning, local for files you would never upload — that's the architecture every desktop agent will copy. Hybrid compute isn't a Mac feature. It's the first honest answer to 'where does my data actually go?'

Like most other harnesses, moving on…

cloud for the thinking, local model when the file gets private. that's the split i actually want on a mac

couldnt ai fully run on user's computer? thats a serious question

Did you remove Comet ability to interact with browser tabs? That was exactly the feature that would have made us use it.

Hybrid compute is a practical answer to the local-versus-cloud debate. Keep sensitive context on the device, send bursty work where capacity is cheaper, and make the handoff visible. Privacy is a routing decision, not a slogan.

That's nice, if there's no performance drop them it'll be great for research

just put the down round in the bag bro

the handoff boundary is the interesting part here. would love to see what stays local, what returns to the cloud, and how that is made visible to the user

Hybrid compute buys privacy at the cost of capability — the local-model steps are necessarily weaker and slower than the cloud ones. It's a coherent split only if users accept 'sensitive work gets the worse model,' which is a real product constraint.