Perplexity Portable Computer Runs Perplexity Computer on Hardware You Own

Perplexity Portable Computer is a local version of the Perplexity Computer agent that runs on your own NVIDIA hardware.

At Layer3Labs, we build and operate AI systems inside other people's businesses. One of the first questions clients ask about local agents is simple: which parts of the task actually stay on the machine?

Announced on 25 August 2026 and built with NVIDIA, Portable Computer keeps model inference, planning, and private-document processing local. Its reasoning never leaves the machine, though web searches and connector calls still go out.

With your permission, it can also consult a frontier advisor model in the cloud. That model only returns text guidance, it cannot access your local files or tools.

Portable Computer is included with Pro, Max, Enterprise Pro, and Enterprise Max subscriptions rather than sold separately. At launch, it supports Linux only, with Windows support planned for September 2026.

What Perplexity Portable Computer Is

Perplexity Portable Computer is the local-first edition of Perplexity Computer, the agent Perplexity already runs as a cloud product.

The difference is location. Instead of sending a task to Perplexity's servers and waiting for an answer, Portable Computer executes the whole agent loop on a machine you own. The contents of the files it reads never travel.

Perplexity built it with NVIDIA and announced it on 25 August 2026. It is not sold on its own, and it is not a separate app you buy once.

Portable Computer unlocks with a qualifying Perplexity subscription and then runs on either an NVIDIA DGX Spark or a Linux PC with an NVIDIA RTX graphics card.

• Model inference, served locally through vLLM, so prompts and documents stay on the machine
• The orchestrator, planner, and tool router, which decide each next step without a round trip to the cloud
• The scheduler and durable task queue, which hold long-running jobs on your hardware
• A local search index over your own files, so retrieval does not require uploading them
• Private-document processing and personally identifiable information (PII) classification, both handled on-device

What Runs On-Device and What Leaves It

Portable Computer keeps every reasoning and file-handling step on the device, and sends out only web searches and connector calls.

Model inference runs locally through vLLM, the open-source serving engine Perplexity uses on the device. The planner chooses the next action and the tool router picks each tool, both on your hardware. The scheduler and the durable task queue run there too, so a long job does not stall when Perplexity's cloud is busy.

Private-document processing and PII classification are local too. A contract or a personnel file is read by a model on your own machine, not uploaded to be read somewhere else.

Two things do leave. Web searches go out to the open internet, and connector calls reach whichever service the connector points at.

There is a third, and it is opt-in. With your permission, Portable Computer can escalate a hard problem to a frontier advisor model in the cloud, and that advisor returns text guidance only. It cannot read your local files or run your local tools, so the escalation shares the question rather than the data behind it.

• Stays local: inference, orchestrator, planner, tool router, scheduler, task queue, search index, document processing, PII classification
• Leaves the device: web searches
• Leaves the device: connector calls to Google Drive, Gmail, Slack, and GitHub
• Leaves the device only with permission: an advisor call to a frontier cloud model, returning text guidance and nothing executable

The Two 27B Models at Launch

Portable Computer ships with two 27-billion-parameter models: Qwen 3.8 27B and PPLX 27B.

PPLX 27B is Perplexity's own post-trained version of the Qwen model. Post-training adapts a general model to the agent work Portable Computer does, so the two share the same base model rather than competing on different architectures.

A third model, Nemotron 3.5 Lightning, is listed as coming rather than shipped. Anyone planning around it should treat it as unreleased until Perplexity says otherwise.

Both models advertise a 260,000-token context window. Reporting on the release puts real degradation past roughly 100,000 tokens, so a 250,000-token contract set is unlikely to be handled as well as the advertised figure suggests.

The practical answer to that is chunking. Split a long corpus into pieces that sit under the degradation point. Let the local search index find the right piece instead of pushing a whole archive into one prompt.

• Qwen 3.8 27B, shipped at launch
• PPLX 27B, Perplexity's post-trained variant of the Qwen model, shipped at launch
• Nemotron 3.5 Lightning, listed as coming and not available at launch
• 260,000-token advertised context, with reported degradation past roughly 100,000 tokens

Portable Computer Needs a DGX Spark or a 24GB RTX GPU

Portable Computer runs on an NVIDIA DGX Spark or on a Linux PC with an NVIDIA RTX graphics card.

The DGX Spark uses NVIDIA's Grace Blackwell GB10 chip and was listed at $4,699 on NVIDIA's marketplace as of 27 August 2026. Hardware listings move, so verify current pricing on NVIDIA's DGX Spark product page before you budget for one.

The build-it-yourself route has a floor: 24GB of video memory (VRAM), which means an RTX 3090 or newer. 32GB is cited as the official recommendation. A 24GB card meets the minimum with no headroom above it.

That floor excludes most consumer PCs, and Perplexity lists it as a limitation rather than leaving buyers to find out. If your desktop card sits under 24GB, the choice is a new card, a DGX Spark, or a different local runner such as Ollama.

Multiple DGX Spark units can be linked over shared memory in groups of 2, 4, or 8. That is the path to running larger models than a single box holds. A team that wants one shared local agent, rather than one per desk, buys more units instead of a bigger card.

There is no Apple silicon support, and none on the roadmap. Perplexity has said its focus stays on NVIDIA hardware, so a Mac is not a route to Portable Computer at all.

• NVIDIA DGX Spark (Grace Blackwell GB10), listed at $4,699 as of 27 August 2026
• Or a Linux PC with an NVIDIA RTX GPU
• 24GB VRAM minimum (RTX 3090 or newer), with 32GB cited as the official recommendation
• 2, 4, or 8 DGX Spark units can be linked over shared memory for larger models
• No Apple silicon support, and none announced

Which Platforms and Plans Portable Computer Needs

Linux is the only platform supported at launch, and Perplexity has stated Windows support for September 2026.

That single-platform start narrows the audience sharply in the first weeks. A team with NVIDIA workstations running Windows has the hardware and still has to wait.

Access is tied to a subscription rather than sold separately. Pro, Max, Enterprise Pro, and Enterprise Max subscribers get Portable Computer as part of the plan they already pay for.

There is no standalone Portable Computer price, because the cost is the subscription plus the hardware. Subscription rates change, so check Perplexity's own pricing page for current figures, and see the Perplexity Portable Computer pricing breakdown for how the two halves add up.

• Linux at launch
• Windows support stated for September 2026
• Included with Pro, Max, Enterprise Pro, and Enterprise Max subscriptions
• Not available as a standalone purchase

The Four Connectors Portable Computer Ships With

Portable Computer connects to four services at launch: Google Drive, Gmail, Slack, and GitHub.

A connector call sends a query out to that service and brings the result back for the local model to work on. The reasoning stays on your machine while the retrieval does not.

That split is worth understanding before a compliance review. A team that cannot let contract text reach a third-party model may still have to account for the agent querying a cloud drive that stores those contracts.

For the day-to-day version of this, the Portable Computer setup guide walks through connecting a source and running a first task.

• Google Drive, for documents the agent reads and reasons over locally
• Gmail, for pulling message context into a local task
• Slack, for channel and thread content
• GitHub, for repository content

Perplexity's Reported Benchmark Numbers

Every benchmark figure for Portable Computer so far is Perplexity's own reported number, not independent testing.

On the Local Knowledge Work Bench, a 53-task suite, Perplexity reports 82.6% against 77.6% for Pi, the rival agent harness it benchmarked Portable Computer against.

On BrowseComp, a web-research benchmark, Perplexity reports 66.7%, using 51% less wall time and 70% fewer tokens than Pi. Lower token use shows up on a bill, because a task that finishes with fewer tokens costs less to run at volume.

On Terminal Bench 2.1, a coding benchmark, Perplexity reports 59.6% running fully locally and 73.0% with the cloud advisor enabled, at roughly $0.415 per task.

The 13.4-point distance between 59.6% and 73.0% is the clearest published measure of what the local models give up on hard work. It is also the case for leaving the advisor switched on. Treat all of these as vendor claims until someone outside Perplexity reproduces them.

• Local Knowledge Work Bench, 53 tasks: 82.6% against 77.6% for the Pi comparison harness, Perplexity's reported figure
• BrowseComp web research: 66.7%, with 51% less wall time and 70% fewer tokens than Pi, Perplexity's reported figure
• Terminal Bench 2.1 coding, local only: 59.6%, Perplexity's reported figure
• Terminal Bench 2.1 coding, with cloud advisor: 73.0% at roughly $0.415 per task, Perplexity's reported figure

The Stated Limits of Portable Computer

Perplexity publishes four limitations, and each one decides whether Portable Computer fits a given team.

None of the four is a build defect. They are the cost of moving a frontier-style agent onto a machine that sits under a desk. Three of the four have a workaround that costs money rather than a redesign.

The reasoning ceiling ends the most conversations. If the daily workload is hard multi-step analysis rather than document work, a 27B model on local hardware is the wrong tool, and Perplexity's own Terminal Bench figures say so.

For a fuller side-by-side against the usual local-model stack, see Portable Computer vs Ollama.

• Compact models trail frontier models meaningfully on hard reasoning. Route those tasks to the opt-in cloud advisor or keep a cloud model for them.
• The 24GB VRAM requirement excludes most consumer PCs. The fix is a 24GB or larger RTX card, or a DGX Spark.
• Advertised context of 260,000 tokens degrades past roughly 100,000. Chunk long document sets and lean on the local search index instead of one giant prompt.
• Apple silicon is unsupported with nothing announced. Mac users need a different local runner, such as Ollama or LM Studio.

Portable Computer vs the Cloud Perplexity Computer

The two products run the same agent idea in different places, and the choice comes down to whether you have NVIDIA hardware and a reason to keep files local.

Perplexity has not published a feature-by-feature split between the two, so anything beyond the rows above is not settled yet. Our Portable Computer vs Perplexity Computer comparison tracks what has been published as it lands.

What Portable Computer Changes for Teams Running Local AI

A local-first agent changes the compliance conversation more than it changes output quality. In the workflow automations we run for small and mid-sized business (SMB) teams, the blocker is rarely how smart the model is. It is a policy saying client files cannot be sent to a third-party model.

Portable Computer answers that policy for the reasoning step and leaves it open for the retrieval step, because connector calls still reach Google Drive and Gmail. A review that only asks where inference happens gets a clean answer. A review that asks where the documents are stored does not.

Across the routine automation we build and operate on our own portfolio of sites, the pattern that holds up is split routing. Cheap local steps handle bulk classification and drafting, and only the few real judgment calls go to a larger model. Portable Computer is built for exactly that shape, with the opt-in advisor available for the steps a 27B model cannot handle.

Who should skip it: anyone on a Mac, anyone whose GPU has less than 24GB of VRAM, and any team whose core work is hard multi-step reasoning. Perplexity's own reported Terminal Bench 2.1 result puts local-only performance at 59.6% against 73.0% with the cloud advisor. That gap is the whole argument against a local-only setup for reasoning-heavy work.

Three things would change that. Windows support in September 2026 widens the eligible hardware pool, and Nemotron 3.5 Lightning shipping could lift the local ceiling. Independent benchmarks would also replace vendor claims with something checkable. Apple silicon support would change it most, and Perplexity has said that is not planned.

Before committing to Perplexity Portable Computer, price a DGX Spark against a year of cloud API calls for the same workload with our local AI hardware calculator.