How to Use Perplexity Portable Computer on Your Own Hardware

Before you install Perplexity Portable Computer, there are four things to check: your subscription tier, graphics card, video memory, and operating system.

We haven't run Portable Computer ourselves. But the local-model rollouts we manage for clients often hit the same roadblock: hardware requirements that no one confirmed before the project began.

Perplexity announced Portable Computer on 25 August 2026 as a local-first version of its Perplexity Computer agent, built with NVIDIA. The requirements below are based on what Perplexity shared at launch. Before buying anything, or setting aside your Saturday, confirm each one in Perplexity's documentation.

Confirm Your Subscription Tier Unlocks Portable Computer

Portable Computer unlocks with Pro, Max, Enterprise Pro, and Enterprise Max subscriptions, and it is not sold as a separate purchase.

Access comes from the subscription rather than from the machine. A qualifying graphics card on an account outside those four tiers gets you nothing to install.

On Enterprise Pro and Enterprise Max the tier sits on a company account, so the person who can confirm it is often an admin rather than you. Ask before you order hardware, because the hardware is the expensive half of this and the subscription is the half that can block it.

Confirm which tier your account is on through Perplexity's pricing page, and treat any figure you find there as current only on the day you read it.

• Pro, Max, Enterprise Pro, Enterprise Max: Portable Computer is included
• Any other tier: no access, whatever hardware you own
• Not a separate purchase, so there is no one-off unlock to buy
• Enterprise tiers are administered centrally, so check with whoever owns the account

Confirm the GPU and the Video Memory Before You Buy Anything

Portable Computer runs on an NVIDIA DGX Spark, or on a Linux PC with an NVIDIA RTX graphics processing unit (GPU).

The reported floor is 24GB of video memory, usually written as VRAM, which in practice means an RTX 3090 or newer. Perplexity cites 32GB as its recommendation. A 24GB card qualifies on paper. The extra 8GB leaves room for the model plus everything the agent holds alongside it.

Perplexity lists the 24GB requirement as a limitation of its own, because it rules out most consumer PCs. A gaming laptop with 8GB or 12GB of VRAM does not become eligible through settings.

The DGX Spark is NVIDIA's Grace Blackwell GB10 desktop. Two, four, or eight units can be linked over shared memory, which is how you get past the model sizes a single box can hold.

If you are sizing a machine from scratch, our local AI hardware calculator estimates what a given model needs. For what the hardware and the subscription cost together, see our Portable Computer pricing breakdown.

• DGX Spark (Grace Blackwell GB10), or a Linux PC with an NVIDIA RTX GPU
• 24GB VRAM reported as the floor, meaning RTX 3090 or newer
• 32GB cited as the recommendation, which leaves headroom the floor does not
• Multi-unit DGX Spark setups link in twos, fours, or eights via shared memory

Linux Is the Only Supported Operating System at Launch

Linux is the only operating system Portable Computer supports today.

Perplexity stated Windows support for September 2026. That was a statement made at the 25 August 2026 launch rather than a shipped release. Plan around Linux, and verify the Windows date in Perplexity's documentation before you schedule work against it.

Apple silicon is not supported, and Perplexity has said none is on the roadmap because its focus stays on NVIDIA hardware. A MacBook cannot run Portable Computer at any tier, with any amount of memory.

A Mac team that wants local inference this quarter has two workable options. Buy a separate Linux box with an NVIDIA card, or run a different local runtime such as Ollama, which does support Apple silicon.

• Linux: supported at launch
• Windows: stated for September 2026, unconfirmed as shipped
• macOS and Apple silicon: not supported, and not on the roadmap

Installing Portable Computer and Picking a Model

Two models ship with Portable Computer at launch: Qwen 3.8 27B, and PPLX 27B, which is Perplexity's post-trained version of the Qwen model.

Nemotron 3.5 Lightning is listed as coming rather than shipped, so do not size a project around it yet.

The literal install commands and menu steps come from Perplexity, and they are the part of any launch that changes fastest. Pull them from Perplexity's documentation rather than from a walkthrough written in August.

Context is advertised at 260,000 tokens and reported to degrade past roughly 100,000. Chunk long documents against the 100,000 figure instead of the advertised one, or you will be debugging quality on a job the model was never holding in full.

On the device itself, inference is served through vLLM, alongside the orchestrator, the planner, the tool router, the scheduler, and a durable task queue. Your local search index, private-document processing, and the classifier for personally identifiable information (PII) also run locally, so private documents are processed without leaving your machine.

• Qwen 3.8 27B: the open-weight option at launch
• PPLX 27B: Perplexity's post-trained build of that Qwen model
• Nemotron 3.5 Lightning: announced as coming, unavailable at launch
• 260,000-token advertised context, reported to degrade past about 100,000

Connector Calls and Web Searches Leave the Device

Portable Computer connects to Google Drive, Gmail, Slack, and GitHub, and every one of those connector calls leaves your machine. Web searches leave too.

So the local-first claim covers inference, planning, your search index, and private-document processing. It does not cover the traffic a connector generates when it reads your mail or your repositories.

That distinction decides whether Portable Computer satisfies a compliance rule or breaks one. If the rule is that no query content reaches a third party, a connected Gmail account violates it no matter where the model weights sit.

Across the workflows we have automated for small and mid-sized businesses, we see connector scope decided last, after everything has already been switched on. Enabling all four because they appear on the list is how a setup bought for privacy ends up sending out more than the cloud product it replaced.

Wire one connector, run a week of real work through it, and add the second only when a task fails without it.

• Stays local: model inference, orchestrator, planner, tool router, scheduler, task queue
• Stays local: search index, private-document processing, PII classification
• Leaves the device: web searches and every connector call
• Connectors at launch: Google Drive, Gmail, Slack, GitHub

Decide Whether to Allow Cloud Advisor Escalation

Cloud advisor escalation is optional, requires your permission, and sends text out of the building in exchange for better answers on hard tasks.

The advisor is a frontier model that returns text guidance only. It cannot touch your local files or call your local tools, so an escalation exposes the text of your prompt and nothing else on the machine.

Perplexity's own figures put its local coding score at 59.6% and the same test at 73.0% with the cloud advisor, at approximately $0.415 per task. Perplexity reported both numbers itself, and independent testing has not confirmed them. At volume the per-task price turns a benchmark line into a monthly bill.

Perplexity also reports 82.6% on its 53-task local knowledge-work suite, against 77.6% for the harness it compared with. On web research it reports 66.7%, using 51% less wall time and 70% fewer tokens than that harness. Read all three as Perplexity scoring its own product.

Turn escalation on for coding and research work where the prompt carries no client data, and leave it off for anything covered by a confidentiality obligation. Perplexity states plainly that compact models trail frontier models on hard reasoning. If every task you plan to run is hard reasoning, escalation stops being optional, and the cost per task belongs in your budget.

• Advisor returns text guidance only, with no access to local files or tools
• Vendor-reported coding score: 59.6% local, 73.0% with the advisor
• Vendor-reported cost: approximately $0.415 per escalated task
• Escalation is permission-based, so it stays off until you allow it

Who Should Not Set Up Portable Computer Today

Four groups should not attempt this setup today: Mac owners, Windows-only teams, anyone with under 24GB of VRAM, and accounts outside the four qualifying tiers.

For those readers Ollama is the practical alternative, because it runs on macOS, Windows, and Linux, on CPU or GPU, at no subscription cost. It gives up the agent layer that makes Portable Computer interesting, which is the planner, the scheduler, and the durable task queue. Our Portable Computer versus Ollama comparison sets the two side by side.

Three things would change this answer. Windows shipping as stated in September 2026 opens it to a much larger group. Nemotron 3.5 Lightning arriving gives smaller machines a lighter model to run. A revised VRAM floor below 24GB would bring ordinary gaming PCs into range.

None of those has happened as of 31 August 2026. Until one does, work through the four checks in order before you spend a weekend on how to use Perplexity Portable Computer. Tier first, then the GPU, then its VRAM, then the operating system.

• Not for Apple silicon owners, at any tier
• Not for Windows-only teams until Windows support ships
• Not for GPUs under 24GB of VRAM
• Not for accounts outside Pro, Max, Enterprise Pro, and Enterprise Max