Perplexity Portable Computer vs Perplexity Computer

Perplexity Computer runs on Perplexity's servers. Portable Computer runs on hardware you own. In the client implementations we manage, that choice is usually settled long before anyone starts debating which model is smarter.

Announced on 25 August 2026, Portable Computer is the local-first version of the existing Perplexity Computer agent, built with NVIDIA.

Neither product replaces the other. Both are available, their names are nearly identical, and search results often blur them together. The real question is where you want a task to run, not which version you should upgrade to.

The basic rule is simple: if the work involves documents you cannot share with a third party, Portable Computer is designed for it. If the data is not sensitive and you want the strongest reasoning available, choose Perplexity Computer. The rest of this comparison explains that distinction in more detail.

Before getting into it, a few terms are worth clarifying. Portable Computer serves models through vLLM on your own GPU and includes connectors for Google Drive, Gmail, Slack, and GitHub. Nemotron 3.5 Lightning is listed as coming rather than already shipped, and Portable Computer does not run on Apple silicon.

All scores Perplexity has published for Portable Computer come from Terminal Bench 2.1, BrowseComp, or its own Local Knowledge Work Bench. Each was measured against Pi, the baseline harness used in Perplexity's comparisons.

The two products are not documented equally. Perplexity has published a detailed technical overview of Portable Computer, but much less equivalent information about Perplexity Computer. That is why several cells below say that Perplexity has not published a figure. For current numbers, see Perplexity's pricing page for subscription tiers and NVIDIA's DGX Spark page for hardware pricing.

Perplexity Portable Computer (local) vs. Perplexity Computer (cloud): Side-by-Side

Perplexity Computer and Portable Computer Are Two Different Products

Perplexity Computer and Perplexity Portable Computer are two products that ship at the same time, not two version numbers of one product. Portable Computer arrived on 25 August 2026 as the local-first version of the Perplexity Computer agent, and Perplexity Computer kept running in the cloud. The word Portable carries the entire distinction, which is why one search returns pages about both.

Read as an upgrade question, this goes wrong fast. A buyer assumes Portable Computer supersedes Perplexity Computer and inherits everything it can do. It does not. The models are smaller, the operating-system list is shorter, and the entry hardware bill starts in the four figures.

Read as a routing question, it gets easy. You have one task in front of you, and you are deciding which of Perplexity's two agents should run it. That decision can go a different way for the next task the same afternoon, and running both is a normal outcome rather than a sign you picked wrong.

What Leaves the Device in Each Version

Portable Computer keeps model inference and every planning step on your machine and sends out only web searches and connector calls. Perplexity lists what stays local, and the list is longer than most local-model setups offer.

• Model inference, served through vLLM, so prompts and documents are processed by a GPU you control
• The orchestrator, the planner, the tool router, and the scheduler, so the decision about which tool to call is made on your machine
• The durable task queue, so a long-running job survives on local disk instead of in someone else's session
• The local search index and private-document processing, so your files are indexed where they already sit
• Personally identifiable information (PII) classification, so the screening step never ships the text it is screening

Connectors Call Out in Both Versions

Connector calls leave the device in both versions, and that is the detail readers most often get backwards. Google Drive, Gmail, Slack, and GitHub all work with Portable Computer, and they work the same way with Perplexity Computer. A connector is a network request by definition.

When Portable Computer reads a Gmail thread, the request goes to Google exactly as it would from Perplexity Computer. Local describes where the model thinks. It is not a promise that nothing crosses your network boundary.

There is a practical way to hold both facts at once. If a document must never leave the building, keep it on local disk and let Portable Computer index it there. Do not pull that same document through a connector and assume the local runtime protected it. Web search behaves the same way. A research task that queries the open web has already sent a query out no matter which product you chose.

Model Class and the Reasoning Ceiling

Portable Computer runs 27-billion-parameter models and Perplexity Computer runs frontier models, and that gap sets what each one can finish. Portable Computer shipped with 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 and was not there at launch.

Perplexity states the limitation itself: compact models trail frontier models meaningfully on hard reasoning. Its own reported figures show where the line sits. On a Local Knowledge Work Bench of 53 tasks it reports 82.6% against 77.6% for the Pi harness. On BrowseComp it reports 66.7%, using 51% less wall time and 70% fewer tokens than Pi. On Terminal Bench 2.1 it reports 59.6% running locally and 73.0% with the cloud advisor engaged.

Every one of those numbers is Perplexity's own, and we have not found independent testing to set against them. Read them as a shape rather than a score. Repeatable work holds up locally. The 13.4-point spread on Terminal Bench 2.1 between the local run and the advisor-assisted run is the size of the hard-reasoning gap on that benchmark.

Context length has a similar catch. Portable Computer advertises 260,000 tokens and is reported to degrade past roughly 100,000, so a 200,000-token repository dump is inside the advertised limit and outside the range where quality holds. Split that job before you run it.

What Perplexity Computer and Portable Computer Cost to Run

Portable Computer moves spending from a per-task bill to an up-front hardware bill. An NVIDIA DGX Spark was listed at $4,699 as of 27 August 2026, and that figure moves, so check the current one on NVIDIA's DGX Spark page before you order. The alternative is a Linux PC with an NVIDIA RTX GPU you may already own, which turns the same decision into a much smaller purchase.

After the hardware is paid for, a local task costs electricity. The exception is the cloud advisor. Perplexity reports approximately $0.415 per task for the advisor-assisted Terminal Bench 2.1 runs. A workload that escalates on every job rebuilds a usage bill on top of the hardware you just bought.

Access to Portable Computer comes with a Pro, Max, Enterprise Pro, or Enterprise Max subscription rather than a separate purchase. We do not have a published tier requirement or a per-task price for Perplexity Computer. Subscription prices also change without notice. Confirm both on Perplexity's pricing page before you build a budget around either product.

→ Compare per-token pricing for every AI model

The Hardware and Operating-System Gate

Portable Computer runs on Linux and nothing else today, which disqualifies more buyers than any benchmark result does. Windows support is stated for September 2026. Apple silicon is unsupported with nothing on the roadmap, and Perplexity has said its focus stays on NVIDIA hardware. A Mac-only team cannot run Portable Computer this month or on any published date.

The GPU requirement is the second gate. Reported floor is 24GB of VRAM, meaning an RTX 3090 or newer, with 32GB cited as the official recommendation. That rules out most consumer desktops and every laptop running on an integrated GPU. Check which card sits in the machine you planned to use before anything else.

Teams that need headroom can link DGX Spark units by shared memory in groups of 2, 4, or 8 to run larger models. If you are sizing a build rather than buying a Spark, our local AI hardware calculator estimates the VRAM and GPU tier a given model needs.

When Portable Computer Is the Right Call

Portable Computer wins when the constraint on the task is legal or contractual rather than technical. The work below is what makes the hardware bill worth paying.

• Client files, medical records, or unreleased financials that a contract says may not be processed by a third party
• High-volume repetitive document work, where a per-task cloud bill grows every month and a paid-off GPU does not
• Long-running jobs you want queued on your own disk, using the durable task queue rather than a session you have to keep alive
• An existing Linux workstation with a 24GB or larger NVIDIA RTX card, which drops the cost of trying this to zero
• Search across private documents that you would rather index locally than upload anywhere

When Perplexity Computer Is the Right Call

Perplexity Computer wins whenever the reasoning ceiling matters more than the data boundary. The frontier models behind it do work the 27B local models are documented to struggle with, and nobody has to buy a GPU first.

• Hard novel reasoning, planning, or analysis, where Perplexity's own Terminal Bench 2.1 numbers show the local run trailing the advisor-assisted run
• Work on public data, where the argument for running locally never applies
• A Mac team, or any team without a Linux box and a 24GB NVIDIA GPU, since Portable Computer will not run at all
• Occasional use, where a $4,699 machine is hard to justify
• Tasks that are mostly web research anyway, since the search queries leave your network in either version

Who Should Skip Both

A team with a strict data-residency rule and a Mac-only fleet should not buy either one yet. Portable Computer does not run on Apple silicon, and Perplexity Computer processes the task on Perplexity's servers, so neither satisfies the rule. Wait for the stated September 2026 Windows release and pair it with a Windows workstation carrying a 24GB card. The other route is an open-weight model run locally through something like Ollama.

Anyone whose real problem is one Gmail inbox and a spreadsheet should also stop here. An agent that plans, routes tools, and runs a durable queue is built for multi-step work. A smaller automation finishes that job for far less.

What Would Change This Answer

Four specific changes would move our recommendation, and three of them are already dated or announced.

• Windows support shipping as stated in September 2026, which converts a large group of Windows workstation owners from skip-both into candidates for Portable Computer
• Nemotron 3.5 Lightning arriving and closing part of the reasoning gap, which would reduce how often the paid cloud advisor is needed
• Independent benchmark results appearing, since every figure available today is Perplexity's own and a third-party run could move the picture in either direction
• Apple silicon support, which Perplexity has said is not planned and would reopen the local option for design and executive teams running Macs

How to Route One Real Task This Week

Pick a task you already run by hand and route it deliberately instead of guessing. Four steps settle it.

• Write down what the task touches, then mark every file that a contract or a client would object to sending off your network
• If nothing on that list is restricted, run it on Perplexity Computer first, because it needs no hardware and gives you the quality bar to beat
• If something is restricted, check the machine you would use for Linux and a 24GB or larger NVIDIA GPU before you spend anything
• Time the run and count how often the job escalates to the cloud advisor, since a task that escalates constantly is a task the local models are not sized for

The Verdict

Run the sensitive work on Portable Computer and the hard work on Perplexity Computer. With Portable Computer, model inference, planning, and private-document indexing stay on your machine. That answers the contract question that stops most rollouts. It does so with 27B models that Perplexity itself says trail frontier models on hard reasoning.

Hardware and operating system are the gates, and price is not one of them. Linux only until the stated September 2026 Windows release, no Apple silicon at all, and a 24GB VRAM floor with 32GB recommended. A team that clears those gates and already owns the GPU can test Portable Computer for the cost of an evening. A team that does not clear them is choosing Perplexity Computer whether they wanted to or not.

Do not treat connectors as the dividing line, because Google Drive, Gmail, Slack, and GitHub calls leave the device in both versions. List the files your next task touches. Send that task to Perplexity Computer if none of them are restricted, and to Portable Computer if any of them are.