Portable Computer for Windows went live on 14 September 2026, three weeks after Perplexity first shipped its local agent for Linux and NVIDIA’s DGX Spark. The release puts the same agent harness, orchestrator, scheduler and a post-trained local model inside the existing Perplexity app for Windows, provided the PC carries an NVIDIA GeForce RTX or RTX PRO graphics card with at least 24GB of video memory. Work the agent finishes on the device does not consume Computer credits.

That summary is accurate, and it leaves out the details that decide whether the update matters to you. Perplexity’s own product page says Windows machines get one local model, PPLX 27B, and that Qwen 3.8 27B “is not available on Windows RTX PCs”, while NVIDIA’s launch blog names Qwen 3.8 27B as the model the app sets up. Setup starts inside a different product, Personal Computer for Windows. And the 24GB rule excludes most gaming PCs, never mind office laptops: in Steam’s August 2026 hardware survey, just 1.50% of systems reported one of the three qualifying GeForce cards the survey lists.

We covered the original launch in detail in Perplexity’s Portable Computer launch with Nvidia, including the benchmarks and the credit arithmetic. This article is about what Portable Computer for Windows changes: which GPUs qualify, which model you actually get, how scheduled tasks and local MCP servers work, where the privacy boundary sits, and where the launch-day coverage got the details wrong. For every other local and cloud release, our AI models, tools and releases hub keeps the running list.

What Portable Computer for Windows Actually Adds

perplexity portable computer for windows nvidia rtx gpus b rice cooker with domed lid

Portable Computer is the local edition of Perplexity Computer, the agent that plans and carries out multistep tasks. Perplexity says the Windows release “adds Computer’s local agent harness and orchestrator to compatible Windows PCs”, and that the model, agent harness, orchestrator and scheduler run “entirely on the Windows device”. Files, queries and agent activity stay on the PC unless you approve something leaving it.

The same harness, now on the Windows desktop

The harness is the part Perplexity built around a small model rather than borrowing from a general-purpose agent framework. Its August research post describes a deliberately short system prompt, skills that load and unload on demand, context compaction when a job runs long, and an OS-level sandbox for every tool call. Portable Computer for Windows inherits all of that. What changes is the desktop around it: Windows files, Windows applications and the Microsoft tools most offices already run.

Scheduled tasks are the headline feature

The Windows announcement leads with scheduling. Users can set recurring workflows, such as analyses or file processing, to run on a local model on their own hardware instead of doing them by hand. Perplexity’s worked example is a logistics controller who schedules Portable Computer for Windows to reconcile freight invoices in a permitted folder every morning, match each one against the carrier’s local rate sheets, flag duplicate charges and off-contract rates, and produce an exception queue with a recommended dispute action for each.

“The invoice and rate data never leave the company’s device,” the company says. That sentence is the whole commercial pitch of Portable Computer for Windows in one line: recurring, document-heavy work that a finance or operations team would never paste into a chatbot, running on a timer, on hardware the company controls.

Local MCP servers widen what the agent can touch

The second addition is local Model Context Protocol support. Portable Computer for Windows can connect to other desktop applications through MCP servers running on the same machine, so the local model can use those applications’ tools inside a workflow. Perplexity’s example is a designer connecting it to an installed application to create or revise a file, with the model, the source files and every application interaction staying local.

That is a notable turn. For the Linux launch, Perplexity’s research post explains, the team converted the most-used connectors from MCP servers into compact command-line tools, because MCP tool definitions “consume a substantial share of the context” a small model can use well. Local MCP on Windows reopens that door for desktop software. Whether a 27B model copes with a large MCP tool surface is something to test on your own applications rather than assume.

The agent can also work with documents, spreadsheets, presentations, PDFs, code and images stored in folders you permit, then act on them with tools. The permitted-folder model is worth planning around: point it at a working directory, not at a whole user profile.

Portable Computer for Windows vs the August Linux Launch

perplexity portable computer for windows nvidia rtx gpus c junction box cube with three conduit stubs

The Linux release on 25 August and the Windows release on 14 September share one core, but several details moved. The table sets the two side by side using only what Perplexity and NVIDIA published.

DetailLinux launch, 25 Aug 2026Portable Computer for Windows, 14 Sep 2026
Reference hardwareNVIDIA DGX Spark, with RTX PCs “soon”GeForce RTX and RTX PRO GPUs with 24GB+ VRAM
Operating systemLinuxWindows, inside the Perplexity app
Local modelsPPLX 27B or Qwen 3.8 27B on DGX SparkPPLX 27B only
SetupOne-click setup via the Perplexity appPersonal Computer for Windows first, then local inference in Settings
Desktop application toolsConnectors converted to command-line toolsAdds local MCP servers for desktop applications
Connectors named by PerplexityGoogle Drive, Gmail, Slack, GitHubGmail, Outlook, Slack, GitHub
SchedulerRuns on deviceRuns on device, promoted as the lead feature
PlansPro and Max subscribersPro and Max, individual and enterprise plans
Cloud escalation15+ frontier models, on permission15+ frontier models, on permission

What stayed the same

The permission model did not change. When a task needs to send anything from the device to a cloud service, Portable Computer for Windows asks first. Cloud escalation still reaches Perplexity Search, wide or deep research, and “15+ frontier models” for harder reasoning. On-device work still carries no credit charge, and the sandbox is still mandatory rather than optional.

What quietly changed

Two things narrowed and one widened. On RTX hardware the local model choice is PPLX 27B alone, where the Spark offers two. The setup path now depends on a second Perplexity product being installed first. The hardware base, by contrast, widened from a single desk-sized NVIDIA system to any Windows PC with a qualifying card, which is the change that turns Portable Computer for Windows from a developer curiosity into something an IT team might be asked to approve.

The timing also held. Perplexity said in August that Windows support was “coming soon”, and our launch coverage recorded a September target. Delivering Portable Computer for Windows three weeks later is on schedule, which is worth noting in a category where “soon” often means next year.

Which RTX GPUs Run Portable Computer for Windows

perplexity portable computer for windows nvidia rtx gpus d small garden shed with a closed door

NVIDIA’s rule is one line: Portable Computer “is available for NVIDIA GeForce RTX and RTX PRO GPUs with 24GB or more of VRAM”. Perplexity refers to “supported eligible” GPUs and has not published a model-by-model list, so the table applies the 24GB rule to NVIDIA’s own memory specifications.

GPUForm factorMemoryMeets the 24GB rule?
GeForce RTX 5090Desktop32GB GDDR7Yes
GeForce RTX 4090Desktop24GB GDDR6XYes
GeForce RTX 3090 TiDesktop24GB GDDR6XYes
GeForce RTX 3090Desktop24GB GDDR6XYes
GeForce RTX 5090 Laptop GPULaptop24GB GDDR7On paper, yes
GeForce RTX 5080Desktop16GB GDDR7No
GeForce RTX 5070 TiDesktop16GB GDDR7No
GeForce RTX 4080 SUPER and RTX 4080Desktop16GB GDDR6XNo
GeForce RTX 5080 Laptop GPULaptop16GB GDDR7No
RTX PRO 4000 BlackwellWorkstation24GB GDDR7 with ECCYes
RTX PRO 4500 BlackwellWorkstation32GB GDDR7 with ECCYes
RTX PRO 5000 BlackwellWorkstation48GB or 72GB GDDR7 with ECCYes
RTX PRO 6000 Blackwell Workstation Edition and Max-QWorkstation96GB GDDR7 with ECCYes

The 24GB rule in plain terms

A 27-billion-parameter model has to fit in video memory together with its working context, which is why the line sits at 24GB and why 16GB cards miss out however new they are. The RTX 5080 is two generations newer than the RTX 3090, and the older card qualifies for Portable Computer for Windows while the newer one does not. For anyone buying hardware, the rule is simple: check the memory figure, not the model number.

Laptops are a grey area

NVIDIA’s laptop specifications list the GeForce RTX 5090 Laptop GPU with 24GB of GDDR7, which meets the stated threshold, while the RTX 5080 Laptop GPU carries 16GB and does not. Neither Perplexity nor NVIDIA mentions laptops in the Windows announcements, and both describe support in terms of “compatible” or “supported” PCs. If you plan to run Portable Computer for Windows on a 5090 laptop, treat it as likely rather than promised, and confirm in the app before buying anything.

The newer RTX Spark superchip laptops are a separate question. They use a large pool of unified memory rather than a discrete card with its own VRAM, the same broad design as the DGX Spark that Portable Computer already supports on Linux. Neither company has said whether those Windows machines will count.

RTX PRO workstations are the safer business bet

The RTX PRO line is where business buyers will find the most headroom, from 24GB on the RTX PRO 4000 Blackwell up to 96GB on the RTX PRO 6000 Blackwell, according to NVIDIA’s workstation specifications. NVIDIA also says support for its DGX Station is “expected to come soon”, which would extend Portable Computer beyond PCs and the Spark to NVIDIA’s larger deskside system.

How Few PCs Can Run Portable Computer for Windows

perplexity portable computer for windows nvidia rtx gpus e toy delivery truck with a box body

Steam’s monthly hardware survey is the largest public sample of Windows PC graphics cards. It skews heavily towards gaming hardware, so it almost certainly overstates how common a 24GB card is in a typical office. Even so, the August 2026 figures are sobering for anyone hoping to roll Portable Computer for Windows out widely.

Steam Hardware Survey, August 2026: share of all surveyed systems (bars scaled to the RTX 5080 at 1.71%)
RTX 5080, 16GB, does not qualify 1.71%
RTX 4090, 24GB, qualifies 0.71%
RTX 4080, 16GB, does not qualify 0.67%
RTX 4080 SUPER, 16GB, does not qualify 0.63%
RTX 5090, 32GB, qualifies 0.43%
RTX 3090, 24GB, qualifies 0.36%

1.50% qualify and 3.01% just miss

Add the three qualifying GeForce cards the survey lists, the RTX 4090 at 0.71%, the RTX 5090 at 0.43% and the RTX 3090 at 0.36%, and you get 1.50% of surveyed systems. Add the three high-end 16GB cards in the same chart, the RTX 5080 at 1.71%, the RTX 4080 at 0.67% and the RTX 4080 SUPER at 0.63%, and you get 3.01%. Roughly twice as many Steam users own a premium NVIDIA card that falls short of Portable Computer for Windows as own one that clears the bar.

Qualifying vs near-miss GeForce cards, Steam August 2026 (sums of the figures above)
Premium 16GB cards that do not qualify 3.01%
24GB and 32GB cards that qualify 1.50%

What this means for an office fleet

The survey’s audience buys graphics cards for games, so business laptops and standard office desktops will sit far below even that 1.50%. The realistic base for Portable Computer for Windows is not “Windows users”. It is engineering, design, data science and research workstations that already carry a big card, plus anyone willing to buy one for the purpose.

That is also why the headline reach claims need care. Perplexity’s July launch post said Personal Computer for Windows brings its agent platform to “more than a billion devices globally”. That describes the cloud product, which runs on ordinary PCs. Portable Computer for Windows is the local layer on top, and its hardware gate shrinks the addressable base by orders of magnitude.

One Model, Not Two: PPLX 27B on Portable Computer for Windows

perplexity portable computer for windows nvidia rtx gpus f closed logbook with a strap band

The model detail is where the two launch posts diverge. NVIDIA’s blog says the app “simplifies setup with a local model, such as Qwen 3.8 27B, that is post-trained to work with Perplexity Computer and optimized for NVIDIA RTX GPUs”. Perplexity’s product page is more specific: on Windows, “Available local model: PPLX 27B. Qwen 3.8 27B is not available on Windows RTX PCs.”

PlatformLocal models Perplexity listsNotes
NVIDIA DGX SparkPPLX 27B and Qwen 3.8 27BThe August reference machine
Linux PC with a 24GB+ RTX GPUPPLX 27B“Qwen 3.8 27B is not available on Linux RTX PCs”
Windows PC with a 24GB+ RTX GPUPPLX 27BSet up after Personal Computer for Windows
NVIDIA DGX StationNot yet listedNVIDIA says support is “expected to come soon”
Any supported systemNVIDIA Nemotron 3.5 LightningListed as coming soon
Apple silicon MacNoneNot a supported platform

Why the two descriptions do not really conflict

PPLX 27B is Perplexity’s post-trained version of Qwen 3.8 27B. Read that way, NVIDIA’s phrase “a local model, such as Qwen 3.8 27B, that is post-trained to work with Perplexity Computer” is describing PPLX 27B by its base model. The practical outcome is the same either way: on an RTX PC, Portable Computer for Windows gives you the post-trained model, not the stock Qwen weights, and you cannot switch to the base model for comparison.

PPLX 27B is the stronger of the two anyway

Perplexity trained PPLX 27B in two stages, rejection fine-tuning followed by reinforcement learning, inside synthetic environments built from the kinds of tasks people run in Computer and containing no real user documents. On its 53-task Local Knowledge Work Bench, PPLX 27B scored 85.4% against 82.6% for base Qwen 3.8 27B in the same harness, at the cost of more tokens, 678k against 520k.

Those results come from a DGX Spark, not a Windows RTX PC. Perplexity has not published any Windows-specific accuracy, speed or token figures for Portable Computer for Windows, so the numbers are a reasonable guide to the model’s quality and no guide at all to how a 24GB card will cope with long contexts.

Only one local model runs at a time

The product page adds a constraint that matters for scheduling: “Only one local model runs at a time.” If you plan several recurring jobs, they share one model on one GPU. That is a queue, not parallel capacity, so stagger schedules rather than stacking every job at 9am, and expect a game, a render or a CAD session on the same card to compete for the memory.

Setting Up Portable Computer for Windows Starts With Personal Computer

Perplexity’s Windows instructions are two steps: “Start with Personal Computer for Windows, then set up PPLX 27B for local inference in Settings on a supported PC.” The announcement adds that you download the local model you want “in one click after selecting it in the model dropdown”.

Personal Computer and Portable Computer are different products

Personal Computer for Windows launched on 28 July 2026. It brings Computer’s cloud agents to the Windows desktop, working across local files, Microsoft 365 and the web, and it is available to Pro, Max and Enterprise subscribers on Windows 10 and 11. Portable Computer for Windows is the local-inference layer inside that app. The names are close enough to confuse a procurement request, so the table separates them.

DetailPersonal Computer for WindowsPortable Computer for Windows
Launched28 July 202614 September 2026
Where the model runsPerplexity’s cloudYour RTX GPU
CreditsConsumed by agent workNot consumed by on-device work
HardwareA Windows 10 or 11 PCA Windows PC with a 24GB+ GeForce RTX or RTX PRO GPU
PlansPro, Max and EnterprisePro and Max, individual and enterprise plans
ConnectorsMicrosoft 365 plus 400+ App ConnectorsGmail, Outlook, Slack and GitHub through the local orchestrator, plus local MCP servers
Guard on risky actionsAlerts before sending an email or deleting a fileAsks before sending anything to a cloud service

Pro and Max only

Portable Computer for Windows is available “to Pro and Max subscribers on both individual and enterprise plans”. Pro costs $20 a month and Max $200 a month with 10,000 credits, the figures we reported for Computer in August. There is no free tier for local inference, even though the inference itself runs on a graphics card you paid for. The subscription is the price of the harness, the orchestrator and the model, not of the compute.

What Portable Computer for Windows Saves on Credits

The cost argument is the one Perplexity made in August, now aimed at a much bigger installed base: “When people run work on device, it doesn’t consume Computer credits, making long-running workflows more practical.” The interesting question is how much that is worth once scheduling enters the picture.

The scheduler is where the saving compounds

A one-off task saves a one-off amount; a scheduled task saves it every day. Our effort selector coverage recorded the reference figures: Max at $200 for 10,000 credits, so 2 cents a credit; an eight-minute research task measured at 225.71 credits; and a heavier due-diligence workflow at around 500 credits. The chart runs one such job daily for a 30-day month in the cloud product.

One daily job for 30 days, in credits (bars scaled to 15,000 credits)
500-credit workflow daily: 15,000 credits, $300.00 at 2 cents 15,000
Max plan monthly allowance: 10,000 credits, $200 10,000
225.71-credit research task daily: 6,771.3 credits, $135.43 6,771
The same jobs finished on device 0

One daily 500-credit workflow uses 150% of a Max allowance in a month, and one daily 225.71-credit task uses 67.7% of it. Run either as a scheduled job on Portable Computer for Windows and the on-device portion costs nothing in credits. Two caveats: the credit figures come from Perplexity’s cloud product, and a smaller local model may take more steps to finish the same job, so measure your own workload rather than borrowing ours.

Escalation is not free

Local work is free at the margin; escalated steps are not. Perplexity’s research post measured advisor escalation on Terminal Bench 2.1, an 89-task coding benchmark. The local Qwen 3.8 27B scored 59.6%. Allowed to consult a Claude Opus 5 advisor, it reached 73.0% at an estimated API cost of $0.415 per rollout. Claude Opus 5 working alone in the same harness scored 82.4% at $0.65 per rollout.

Terminal Bench 2.1 score, with estimated API cost per rollout (Perplexity’s evaluation)
Qwen 3.8 27B, local only, no API cost 59.6%
Qwen 3.8 27B with a Claude Opus 5 advisor, $0.415 73.0%
Claude Opus 5 alone, $0.65 82.4%

The arithmetic behind “three-fifths”

The gain from escalation is 13.4 points (73.0 minus 59.6; Perplexity quotes 13.5, presumably from unrounded scores). That recovers 58.8% of the 22.8-point gap to Opus 5 alone, for 63.8% of Opus 5’s per-rollout cost ($0.415 divided by $0.65). Perplexity summarises it as “roughly three-fifths of the gap to the frontier at about two-thirds of the frontier’s cost”, and the user decides whether advisor calls are allowed at all and whether each one needs manual approval.

Privacy Boundaries in Portable Computer for Windows

Privacy is the other half of the pitch, and Perplexity’s documentation is unusually specific about where the boundary sits. That makes Portable Computer for Windows easier to assess than most agent products, provided you read the research post rather than the launch blog.

Permission before anything leaves the PC

“When a task needs to send something from the device to the cloud, Portable first asks for the user’s permission before moving forward.” Perplexity’s pull-request example shows the pattern: the agent reviews open pull requests against the local codebase and groups them as ready to merge, needs review, blocked or stale, and only after the engineer approves the summary does it post to the team’s Slack channel and tag the owner of each item.

A PII classifier runs before advisor calls

For escalation to a frontier model, the research post says the harness selects the relevant context, “applies a PII classifier to flag sensitive information, and shows the user what would leave the device”. The advisor receives only the approved context and returns text guidance. It gets no direct access to the device’s files, tools or conversations.

The sandbox refuses to degrade

Tools run in an OS-level sandbox that restricts processes, filesystem paths and network access by policy, which NVIDIA’s post calls Perplexity’s “proprietary SPACE sandbox”. If the sandbox is unavailable, “the harness disables itself before any tool calls rather than degrading to unsandboxed execution”. Perplexity contrasts that with the open-source harnesses Pi and Hermes, which run commands with the user’s own permissions by default.

For anyone writing a data protection assessment of Portable Computer for Windows, that fail-closed behaviour is the most useful sentence in the documentation. The questions left open are the connectors, web search and Slack posts, each of which crosses the boundary by design once approved, and whether approvals are logged anywhere an auditor can see.

Where the Coverage of Portable Computer for Windows Went Wrong

Launch-day coverage leaned heavily on the NVIDIA post, and several details did not survive the trip. The table checks each claim against Perplexity’s own pages.

ClaimWhere it appearedWhat Perplexity’s pages say
The app sets up Qwen 3.8 27BNVIDIA blogWindows RTX PCs get PPLX 27B; “Qwen 3.8 27B is not available on Windows RTX PCs”
The models include “Qwen3.5 27B”AlphaSignalThe base model is Qwen 3.8 27B
Connectors include OneDrive, Word and Google DriveNVIDIA blogThe Windows post names Gmail, Outlook, Slack and GitHub for the local orchestrator
A list of qualifying GeForce and RTX PRO cardsPC GuideNo model list is published; the 24GB rule also covers the RTX 5090 Laptop GPU and the RTX PRO 4000 and 5000 Blackwell
Linux support came first, on 25 AugustPC GuideCorrect
On-device work “consumes zero account credits”AlphaSignalCorrect for on-device work; escalation and the subscription still cost money

The connector list is a real ambiguity

NVIDIA’s list, Outlook, OneDrive, Word, Google Drive, Gmail, Slack and GitHub, may describe everything the Perplexity app connects to, while Perplexity’s shorter list describes what works “through the local orchestrator”. The difference matters because a connector that routes through Perplexity’s cloud is a data flow your privacy review has to cover. Ask which path each connector takes before you enable it in Portable Computer for Windows.

The benchmarks are still from a Spark

Every accuracy number in circulation, from 85.4% on the Local Knowledge Work Bench to 65.1% on ParseBench-100, was measured on a DGX Spark with 128GB of unified memory. A 24GB RTX card has far less memory for context, and Perplexity has not said whether Windows runs use the same context settings. Treat the numbers as indicative for Portable Computer for Windows, not proven.

Who Should Install Portable Computer for Windows Now

The release is genuinely useful for a narrow group and premature for everyone else. The split is mostly about hardware you already own and data you cannot send out.

Strong fits

Teams that already have 24GB workstations on the desk, in CAD, visual effects, data science or research, are the obvious first users, because the hardware cost is sunk. Finance, legal and operations teams with recurring document work, like Perplexity’s freight-invoice example, are the second group. Heavy Computer users whose scheduled jobs would otherwise burn through a Max allowance are the third.

Wait for now

Anyone on 16GB cards, standard business laptops or a Mac should wait; Perplexity’s approach for the Mac is a different design, covered in our piece on Hybrid mode for Mac. So should anyone whose workload is mostly hard reasoning that a 27B model will escalate anyway, and any organisation that needs a clear position on connector data flows before switching agents on.

A two-week pilot

Pick one workflow you can measure. Run it in the cloud product for a week and log the credits. Run it as a scheduled task on Portable Computer for Windows for a week and log escalations, failures and any contention for the GPU. Compare the two logs before buying hardware. It is the same discipline that makes any AI strategy survive contact with real work, and it keeps an agent pilot honest about what autonomous AI agents can finish without help.

Frequently Asked Questions About Portable Computer for Windows

Is Portable Computer for Windows free?

No. It requires a Perplexity Pro or Max subscription on an individual or enterprise plan. Work finished on the device consumes no Computer credits, but steps you approve for cloud escalation do, and you supply the graphics card.

What GPU do I need?

An NVIDIA GeForce RTX or RTX PRO GPU with at least 24GB of VRAM. On the desktop that means an RTX 3090, 3090 Ti, 4090 or 5090, or an RTX PRO card with 24GB or more. The RTX 5080, 5070 Ti and 4080 have 16GB and do not qualify.

Does Portable Computer for Windows run on a laptop?

The GeForce RTX 5090 Laptop GPU has 24GB of VRAM, which meets NVIDIA’s stated rule, but neither company has confirmed laptop support. Check in the app before buying a laptop for this purpose.

Which model does it use?

PPLX 27B, Perplexity’s post-trained version of Qwen 3.8 27B. Perplexity’s product page says the base Qwen 3.8 27B is not available on Windows RTX PCs, and only one local model runs at a time.

Does my data leave the PC?

Not by default. The model, files, queries and agent activity stay local. Anything sent to a cloud service, including web research, advisor calls and posts to apps such as Slack, requires your permission first, and advisor calls pass through a PII classifier.

How is it different from Personal Computer for Windows?

Personal Computer for Windows runs Perplexity’s cloud agents on any Windows 10 or 11 PC and consumes credits. Portable Computer for Windows adds local inference on a 24GB RTX GPU inside that app, so on-device work consumes none.

References