Where the 36GB number comes from — and what "fully local" does not mean
Two questions keep coming up after the launch, and our requirements page answers neither precisely enough: why 36GB, and does fully local mean disconnected? Both have a sourced answer.
1. The arithmetic behind 36GB
Perplexity's own engineering write-up on its Lily inference engine gives the key figure: the Qwen3.6-35B-A3B model, at 4-bit groupwise quantisation, drops from roughly 70GB of bfloat16 weights to a 19.4GB checkpoint. That is the model alone, before anything else runs.
36GB minus 19.4GB leaves about 16.6GB — and that headroom is not spare. It has to hold macOS, the Perplexity app, the isolated sandbox that runs code and tools, and the KV cache, which grows as the conversation context grows, plus whatever else you have open while a task runs. That last sentence is our arithmetic, not Perplexity's: it explains why a recommendation that looks high next to the checkpoint alone is in fact tight, not generous.
One correction to be precise about: a 16.4GB peak figure sometimes quoted alongside this story belongs to a 27B build measured on a 36GB MacBook Pro — not to the 35B-A3B checkpoint. We do not use it as evidence for the 35B model. The 19.4GB checkpoint figure above is the one Perplexity published for Qwen3.6-35B-A3B.
Perplexity's wording matters here too: 36GB is the recommended figure for fully local mode, not a hard lock. Our checker still treats below-36GB as "does not clear the fully local line" — because buying below the recommendation and hoping is not a plan — but the honest label is "recommended", and this page uses it.
2. Fully local describes where the work runs, not whether a network is ever involved
In fully local mode the model and the agent directing the task run on your Mac, and files only reach the cloud when you approve it. Work on files and apps already on the Mac — reading recent mail and messages, spotting deadlines, writing an action list in Notes — is the offline-capable core.
But "the AI runs locally" is not the same as "the task never touches the internet". A task that needs current information, a browser, or a connected app still needs those services: sending a message or a meeting invitation goes through that app's service, and Perplexity says the local agent can, with your permission, escalate to one of more than 15 frontier cloud models for harder reasoning. So the right mental model is: local by default, cloud only by permission, and only for the step that actually needs it — not "disconnected".
That distinction also keeps Hybrid Compute honest. Hybrid, available since September 2026, starts at 24GB (32GB recommended) and is the realistic option for 24–32GB Macs — but it is not fully local, because the cloud still does the reasoning. See the Hybrid guide for that trade-off.
Frequently asked questions
Why does Perplexity Computer local mode recommend 36GB of memory?
Because the 4-bit Qwen3.6-35B-A3B checkpoint alone is 19.4GB, leaving roughly 16.6GB of a 36GB machine for macOS, the app, the sandbox and the growing context cache.
Does fully local mean Perplexity Computer works with no internet at all?
No. It means the model and agent run on your Mac and files stay local unless you approve otherwise. Tasks needing current information, a browser, or a connected app's service — such as sending a message or meeting invitation — still need a network connection and your permission.
What should a 24–32GB Mac owner do instead?
Use Hybrid Compute (24GB minimum, 32GB recommended). It runs on the machine you already own, but it is not fully local: cloud models still handle reasoning.