Perplexity's Computer agent can now run fully offline on Apple Silicon Macs

Perplexity announced today that Computer, its AI agent for Mac, can now run entirely on-device on Apple Silicon hardware — planning tasks, reasoning through multi-step work, and processing local files without sending any data to the cloud, according to the company's own announcement.

From cloud-assisted to fully local

The fully local mode is the third stage in a progression Perplexity has built out over several months. Personal Computer, which let the agent work across local files and native Mac apps, shipped in May. Hybrid Compute followed in September, pairing cloud-based reasoning models with local models that handled private file processing, coordinating the split automatically. Today's release removes the cloud dependency entirely: the local orchestrator runs Qwen 3.6 or PPLX Pro — Perplexity's own post-trained version of Qwen — directly on the Mac's silicon, handling planning, tool use, and task execution without a network round-trip.

What it can actually do on-device

With permission, the fully local agent can read a project's to-do list in Notes, check a teammate's budget request in Messages, pull planned-versus-actual spending from a Numbers spreadsheet, and assemble a status update with a calendar invite for next steps — the kind of multi-app, multi-step workflow that previously required routing data through Perplexity's cloud infrastructure. Code and tool execution run inside isolated sandbox environments with controlled access to files and connected apps, and users retain the choice to authorize cloud escalation for tasks that genuinely need current information, browser access, or connected cloud apps — routing to one of more than 15 frontier models when local reasoning isn't sufficient.

The hardware requirement is the real catch

Fully local operation requires an Apple Silicon Mac with at least 36GB of unified memory, a bar that excludes the base configurations of most current MacBook Air and entry MacBook Pro models. That threshold reflects a broader pattern in on-device AI: Apple's unified memory architecture lets the CPU, GPU, and Neural Engine share a single memory pool rather than being bottlenecked by discrete GPU VRAM, which is what makes running a capable local model practical at all — but matching cloud-grade reasoning still takes considerably more memory than a typical consumer configuration ships with.

Why the local-first framing matters

The pitch is control, not just privacy in the abstract: for a portfolio analysis built on private brokerage holdings and downloaded market data, or any workflow touching financial, health, or otherwise sensitive personal information, local execution means the task's planning, intermediate reasoning, and output never leave the device. That distinction matters more as AI agents are granted broader access to personal files and third-party apps — the same expanding surface area that has pushed competitors like Zoom to build AI governance and activity-auditing tools into their own platforms this month. Perplexity is betting that a meaningful share of users will choose slower, locally-bounded AI work over faster cloud reasoning specifically because the data boundary is the feature.