Portable Computer is Perplexity’s answer to the single loudest complaint about agentic AI: the meter. Announced on 25 August 2026 and built in close partnership with Nvidia, it takes the agentic Computer product that has been running in Perplexity’s cloud since February and packages the whole thing — model, inference engine, agent harness, tools, connectors and sandbox — to run on hardware you already own. Work that finishes on your own machine consumes no billing credits at all.
That is a bigger shift than it sounds. Most “local AI” so far has meant a chat window talking to a quantised model, useful for drafting and little else. This is a full agent loop: it reads your files, parses PDFs, runs shell commands, takes dictation, browses the web and executes code in an isolated sandbox — on your GPU, with your documents never leaving the box unless you say so. Perplexity has been signalling this direction for months; in August its engineers were spotted testing dynamic compute splitting between local and cloud models, and Portable Computer is that experiment shipping.
The catch is hardware. This is a Linux-first release that wants an Nvidia GPU with at least 24GB of VRAM, which rules out the overwhelming majority of laptops in the world. Below we set out exactly what the product does, what it needs, which models it runs, what Perplexity’s own benchmark numbers show, how the escalation-to-cloud permission flow works, and — most usefully for anyone budgeting — what “zero token costs” does and does not cover. If you already track this space through our AI models, tools and releases hub, this is the release that changes the cost maths.
Table of contents
- What Portable Computer Actually Is
- What Portable Computer Runs Locally And What It Does Not
- Portable Computer Hardware Requirements
- The Models Behind Portable Computer
- What The Portable Computer Benchmarks Show
- What Zero Token Costs Means For Portable Computer
- Why Nvidia Is On The Other Side Of This Deal
- Who Should Try Portable Computer Now
- The Limits Of Portable Computer Today
- Your Next Steps With Portable Computer
- Frequently Asked Questions
- References
What Portable Computer Actually Is
Perplexity Computer launched on 25 February 2026 as a cloud-hosted agentic workspace: you give it a job, it plans, uses tools, browses, writes files and reports back. Portable Computer is the same product with the compute relocated. Perplexity’s VP of engineering for infrastructure and enterprise described the port as bringing “the exact same UI to a fully local app”, carrying “the entirety of the agent harness” with it — not a cut-down companion mode.
The full stack ships to your machine
What gets installed is not just weights. Portable Computer bundles the local model, the inference engine that serves it, the agent tools, the application connectors and an operating-system-level sandbox in which generated code is executed. That last component matters more than it reads: an agent that can run shell commands on your workstation is only safe if there is a boundary around it, and the sandbox is that boundary.
Everything starts on the device
The default is local. Every task begins on your hardware, and your conversation history, your files and the model itself stay there. This is the inversion of the normal arrangement, where a local option is something you opt into for privacy at a cost in capability. Here the cloud is the exception you approve rather than the default you tolerate.
What it can actually touch
The feature list is broad for a first release. Portable Computer can work with files stored on the machine, read PDFs, run shell commands, accept dictation and search the web. It connects to Google Drive, Gmail, Outlook, Slack and GitHub, so results can be pushed into the places work already lives. That connector list is the difference between a clever demo and something a team uses on a Tuesday. Reading a scanned invoice properly leans on computer vision as much as on language modelling, which is why the document-parsing scores further down deserve more attention than the headline ones.
What Portable Computer Runs Locally And What It Does Not
The honest framing is hybrid, not local-only. Perplexity is clear that compact models still trail frontier models on genuinely hard reasoning, so the product is designed around a permission boundary rather than a wall.
| Element | Where it lives by default | Can it leave the device? |
|---|---|---|
| Model weights and inference | Your GPU | No |
| Conversation history | Your machine | No |
| Local files and PDFs | Your machine | Only with approval |
| Code execution | Local sandbox | No |
| A single hard reasoning step | Local model first | Yes, per-step, on permission |
| Web search results | Fetched from the internet | Yes, by definition |
The escalation flow is per-step, not per-task
When the local model hits a step it cannot handle well, Portable Computer asks before sending that one step to a more powerful frontier model in the cloud. The answer comes back and the workflow continues locally. That granularity is the design decision worth studying: it is not “run this whole job in the cloud”, it is “let a specialist look at this one paragraph”. Files and other sensitive material can remain on the device throughout.
Context is managed, not infinite
Long-running agent work eats context. The harness compacts older context as a task stretches out and loads specialised skills on demand rather than holding everything in the window. Perplexity’s own documentation is candid that the launch Qwen model offers a 260K-token context window but begins to struggle beyond roughly 100K tokens — a useful number to hold in mind before pointing Portable Computer at a 400-page contract set.
Portable Computer Hardware Requirements
This is where enthusiasm meets a spec sheet. The requirements are not extreme by workstation standards and are completely out of reach for a typical business laptop.
| Requirement | What Perplexity specifies |
|---|---|
| Operating system | Nvidia DGX OS or Ubuntu |
| Architecture | ARM or x64 |
| GPU | Nvidia RTX, or a DGX Spark |
| Minimum VRAM | 24GB |
| Windows | Scheduled for September 2026 |
| Apple silicon | Not on the roadmap |
| Subscription | Pro, Max, Enterprise Pro or Enterprise Max |
The 24GB floor is the real gate
A 24GB VRAM requirement means a GeForce RTX 3090 or newer in practice. That excludes almost every consumer laptop and a large share of desktops, including plenty of machines their owners would describe as powerful. If your fleet is standard business hardware, Portable Computer is not a fleet-wide rollout this quarter; it is a workstation purchase.
DGX Spark is the reference machine
The headline platform is Nvidia’s DGX Spark, a desk-sized system built on the GB10 Grace Blackwell superchip with a 20-core Arm CPU, 6,144 CUDA cores, 128GB of coherent unified memory at 273GB/s and roughly one petaFLOP of FP4 compute. It carries a $4,699 US Founders Edition price after a February 2026 increase from $3,999, attributed to memory supply. Multiple Sparks can be connected over shared memory to hold larger frontier-class models, which is how the platform scales past one box.
Bring your own model is supported
Perplexity has not locked the runtime. Users can plug in their own models and inference servers, which matters for anyone who has already standardised on a particular open-weight model or serving stack and does not want a second one on the same workstation.
The Models Behind Portable Computer
Three models define the launch line-up, and the distinction between them is worth understanding before you judge any benchmark.
| Model | Origin | Status at launch |
|---|---|---|
| Qwen 3.8 27B | Open-weight, Alibaba lineage | Available |
| PPLX 27B | Perplexity post-trained Qwen variant | Available |
| Nemotron 3.5 Lightning | Nvidia, 30B mixture-of-experts | Coming soon |
PPLX 27B is the interesting one
PPLX 27B is Perplexity’s own post-trained version of the Qwen model, tuned for the agent harness rather than for chat. It is the highest-scoring local option on Perplexity’s evaluations, and it is the clearest signal of where the company thinks its advantage lies: not in owning a frontier model, but in owning the harness that makes a small one behave.
Nemotron 3.5 Lightning explains the partnership
Nvidia released Nemotron 3.5 Lightning on 11 August 2026 — a 30-billion-parameter mixture-of-experts design with roughly 3 billion parameters active per token, built for the grunt-work layer of agentic systems: tool calls, validation, retrieval, formatting, summarisation and classification. Nvidia claims up to four times faster output and around 30% faster agentic task completion against models in its class. A model designed for exactly the work an agent harness generates, arriving two weeks before a local agent harness ships on Nvidia hardware, is not a coincidence.
What The Portable Computer Benchmarks Show
Perplexity published evaluations against two other local agent products, Pi and Hermes. Read them as vendor-run evaluations, because that is what they are — but the margins are wide enough to be interesting, and the tasks are the right ones.
The Local Knowledge Work Bench runs 53 tasks of the kind a knowledge worker actually delegates. Perplexity’s post-trained model leads it.
Browsing and parsing are where the gap widens
On BrowseComp, which tests web research, the gap is far larger than on general knowledge work — 66.7% against 50.2% for Pi and 43.9% for Hermes. On ParseBench-100, which tests document extraction, the spread is dramatic: 65.1% against 34.6% and 13.9%.
Terminal Bench shows exactly what escalation buys
The most instructive number is not a win. On Terminal Bench 2.1 the local Qwen configuration scores 59.6%. Allow the harness to escalate hard steps to Claude Opus 5 as an adviser and it reaches 73.0% — at a stated cost of roughly $0.415 per task. That is the honest shape of the product: local is competent, escalation is a measurable 13.4-point upgrade, and the upgrade has a price tag you can read before you approve it.
What Zero Token Costs Means For Portable Computer
The phrase is accurate and it is narrower than it sounds. Local execution consumes no billing credits. Escalated steps do. So does the subscription that entitles you to the software in the first place.
The cloud product’s meter is the baseline
Perplexity Computer meters usage in credits, and the burn has been the product’s sore point since launch. Max costs $200 a month and includes 10,000 credits; Pro costs $20 a month with credits bought separately. Auto-refill defaults to $200 a month when the allowance runs dry and is configurable up to a $2,000 cap.
Published hands-on figures give the shape of it: a reviewer’s single eight-minute research task touching eight tools consumed 225.71 credits, a complex due-diligence workflow runs around 500 credits, and one reviewer burned $200 of credits in two days building a website. We covered that arithmetic, and the effort selector Perplexity is testing to reduce it, in detail last week.
| Factor | Cloud Computer | Portable Computer |
|---|---|---|
| Marginal cost of a local task | Credits per tool call | Zero |
| Cost of an escalated step | Credits | Credits, on approval |
| Up-front hardware | None | 24GB GPU or a DGX Spark |
| Where your files sit | Perplexity’s infrastructure | Your machine |
| Ceiling on heavy use | Your credit budget | Your GPU’s throughput |
| Works offline | No | Yes, apart from web steps |
The break-even is a real calculation
Swap a metered bill for a capital purchase and the question becomes arithmetic. At the published rate of roughly 500 credits for a heavyweight workflow, a 10,000-credit Max allowance buys about twenty of them a month. A team running well past that allowance, month after month, on tasks a 27B model can handle is exactly the profile where Portable Computer pays for a workstation quickly. A team running four heavy jobs a month is not.
Zero marginal cost changes behaviour, not just budgets
The subtler effect is on how people use the thing. When every retry costs credits, users hesitate, batch their questions and under-use the agent. When local retries are free, they iterate. That behavioural shift is usually where the value of a local deployment actually shows up, and it is worth factoring into any AI strategy that currently treats agent spend as a hard cap.
Why Nvidia Is On The Other Side Of This Deal
Nvidia’s director of developer technology framed the timing as an inflection point for local AI, noting that quantised hobbyist models had not been “super practical” and that a new generation of open-source alternatives changed that. The commercial logic is plainer still.
Every local agent is a GPU that had to be bought
Cloud inference concentrates demand in hyperscaler data centres. Local inference distributes it across workstations, and each of those workstations needs a 24GB-plus card or a DGX Spark. A flagship software partner giving people a concrete reason to buy that hardware is worth more to Nvidia than any single cloud contract of comparable size.
The relationship runs deeper than a launch
Nvidia has reportedly discussed joining a Perplexity funding round at a valuation above $30 billion, up from $20 billion in September 2025, with Perplexity’s annualised revenue reported above $750 million as of August 2026 — growth attributed largely to Computer. A hardware partner with an equity interest in the software that sells its hardware is a familiar pattern, and it tends to produce tightly optimised products.
Nemotron closes the loop
With Nemotron 3.5 Lightning arriving in the model picker, Nvidia supplies the silicon, the operating system image, the reference machine and one of the models. There are not many software categories where a chip vendor owns that much of the stack.
Who Should Try Portable Computer Now
Not everyone, and that is fine for a first release. The profile that fits is specific.
Strong fits
Regulated and confidentiality-heavy work is the obvious one: legal, clinical, financial and defence teams that cannot send documents to a third-party endpoint but would still like an agent to read them. Research and engineering teams with existing RTX workstations are the second, because the hardware cost is already sunk. Heavy Computer users hitting their credit ceiling every month are the third — for them Portable Computer converts a variable bill into a fixed one.
Poor fits, for now
Anyone on a Mac, anyone on Windows before September, anyone whose team runs on standard business laptops, and anyone whose workload is dominated by genuinely hard reasoning that a 27B model will escalate anyway. In that last case you are paying for a GPU and still paying for cloud steps.
A sensible pilot
Pick one workflow you run repeatedly and can measure — invoice extraction, contract summarisation, weekly competitor research. Run it in the cloud product for a fortnight and log the credits. Run the same workflow on Portable Computer and log the escalations. The comparison answers the buying question in a way no benchmark can, and it is the same discipline that makes any deployment of autonomous AI agents survive contact with real work.
The Limits Of Portable Computer Today
A fair assessment has to name what is missing, because the gaps are not small.
Linux only, for another few weeks
Windows support is scheduled for September 2026, which is close enough to be a scheduling problem rather than a strategic one. Until then the addressable audience is people comfortable running Ubuntu or DGX OS on a workstation — a competent group, but not a large one.
Apple silicon is absent, deliberately
The Mac is home to a great many local-model enthusiasts, and unified memory makes it well suited to the job. Perplexity has said it is very focused on Nvidia hardware right now. Given who co-built the product, that is not surprising, but it removes a large and enthusiastic constituency from the launch.
The benchmarks are Perplexity’s own
Every figure above comes from the vendor. The comparisons look carefully constructed and the margins are wide, but no independent evaluation exists yet. Treat them as a claim to verify on your own tasks, not as settled fact.
Small models are still small models
Compact models trail frontier performance on hard reasoning, which is the entire reason the escalation path exists. The 100K practical context limit on a 260K window is a second, quieter constraint. Portable Computer is very good at the large volume of ordinary work; it is not a frontier model on your desk.
Your Next Steps With Portable Computer
If the profile fits, the next steps are cheap. Check whether any machine you already own clears 24GB of VRAM — a surprising number of design and engineering workstations do. Confirm your subscription tier covers it, since the entitlement starts at Pro. Then run the pilot described above rather than a demo, because the credit log is the only number that will settle the decision internally.
If the profile does not fit, the release still matters. A full agent harness running usefully on a single 24GB card is a marker for where the whole category is going, and the September Windows release will move the addressable hardware base by an order of magnitude. The right posture is to know your numbers before that happens.
Frequently Asked Questions
Does Portable Computer really cost nothing to run?
Local execution consumes no billing credits. You still pay for a Perplexity subscription at Pro, Max, Enterprise Pro or Enterprise Max, you pay for the hardware, and you pay credits for any step you approve for escalation to a cloud model.
Can I use Portable Computer on a Mac?
No. Apple silicon is not on the roadmap, and Perplexity has said it is focused on Nvidia hardware for now.
What is the minimum GPU?
An Nvidia RTX card with at least 24GB of VRAM, which in practice means a GeForce RTX 3090 or newer, or an Nvidia DGX Spark.
Does my data leave the machine?
Not by default. The model, your files and your conversation history stay local. Individual steps go to a cloud frontier model only when you grant permission, and web searches necessarily reach the internet.
Which model should I choose?
PPLX 27B scores highest on Perplexity’s own evaluations because it is post-trained for the agent harness. Qwen 3.8 27B is the open-weight baseline, and Nvidia’s Nemotron 3.5 Lightning is expected shortly.
When does Windows support arrive?
September 2026, according to Perplexity. Until then the product runs on DGX OS or Ubuntu, on ARM or x64 systems.
References
Perplexity Computer Goes Local With Portable Computer
Perplexity’s new local-first AI runs on your PC and asks before using the cloud
NVIDIA Nemotron 3.5 Lightning and NeMo Switchyard Deliver Faster, Smarter, More Efficient Agentic AI
nemotron-3.5-lightning-30b-a3b model card
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.