What hardware does Portable Computer require?

At launch it targets NVIDIA silicon on Linux: the DGX Spark desktop supercomputer built on the GB10 Grace Blackwell Superchip with a 20-core Arm CPU and 128GB of unified memory, RTX PRO workstation GPUs, and GeForce RTX cards with at least 24GB of memory. Windows support for RTX machines is stated as coming later. There is no Apple Silicon or AMD path at launch.

Which models does Portable Computer run locally?

It ships with Qwen 3.8 27B and PPLX 27B, Perplexity's own post-trained 27B model. NVIDIA Nemotron 3.5 Lightning is announced as coming soon. For heavier reasoning the agent can escalate to a cloud model, but only after explicit user permission — the default path keeps files, prompts and work on the local machine.

Is Portable Computer free?

No. It is included for Perplexity Pro and Max subscribers rather than sold separately. What is free is the marginal inference: any step the agent completes locally has zero per-token cost, so a long-running agent loop that would be expensive in the cloud costs only electricity. Cloud escalations still bill normally.

Quick Answer

What Is Perplexity Portable Computer? Local AI Agents

Q: What is Perplexity Portable Computer?

Portable Computer is a local-first version of Perplexity's agentic Computer platform, launched August 25, 2026 in partnership with NVIDIA. It runs the agent harness, the orchestrator and post-trained models directly on your own machine inside an OS-enforced sandbox, instead of executing agent steps in Perplexity's cloud. Local steps cost nothing per token because no tokens leave the device.

Published: August 26, 2026

The Short Answer

Portable Computer is Perplexity’s agentic AI platform rebuilt to run entirely on hardware you own. Launched August 25, 2026 with NVIDIA, it installs an agent harness, an orchestrator and post-trained models onto a local machine, executes agent steps inside an OS-enforced sandbox, and charges nothing per token for work completed locally.

The interesting part is not the models. It is that the harness moved.

What Actually Runs On Your Machine

Most “local AI” products are a model runner: they load weights and serve tokens. Portable Computer ships the whole agent stack:

Component	Where it runs	What it does
Agent harness	Local	Tool calls, file access, execution loop
Orchestrator	Local	Plans and sequences multi-step tasks
Models	Local	Qwen 3.8 27B, PPLX 27B (post-trained)
Sandbox	Local, OS-enforced	Contains what the agent can touch
Cloud reasoning	Remote, opt-in	Only after explicit permission

That distinction matters because the harness is where your data actually lives. An agent that reads twelve files to answer one question sends those twelve files somewhere. If the harness is remote, “somewhere” is a vendor’s infrastructure. If the harness is local, the files never move.

Hardware Requirements, Specifically

Portable Computer is NVIDIA-only at launch, on Linux:

NVIDIA DGX Spark — the reference target. GB10 Grace Blackwell Superchip, 20-core Arm CPU, 128GB unified memory, roughly 1 petaFLOP of FP4 compute. Priced at $4,699 since NVIDIA’s February 23, 2026 increase (partner systems from ASUS, Dell, HP, Lenovo and others land near $4,499).
NVIDIA RTX PRO workstation GPUs.
GeForce RTX with 24GB minimum — this is the accessible tier.
Windows on RTX: stated as coming, not shipping.

The 24GB GeForce floor is the number that decides whether this is relevant to you today. A 27B model at reasonable quantisation fits in 24GB with room for context; below that, it does not.

The Cost Argument

Agentic workloads have an awkward economic shape: one user request fans out into dozens of model calls, most of them cheap reasoning steps that are individually trivial and collectively expensive. That is exactly the shape that punishes per-token billing.

Running those steps locally makes the marginal cost of a step zero. A loop that retries, re-reads a file, or explores five approaches instead of one costs the same as a loop that gets it right first time. That changes what you are willing to let an agent attempt — arguably more than any benchmark score does.

The honest caveat: you paid for the hardware up front. A $4,699 DGX Spark is roughly 21,000 GPT-5.6 Sol calls at a 30K-in/5K-out shape (~$0.22 each). If your actual agent usage is a few hundred tasks a month, the cloud is cheaper and will stay cheaper. The break-even favours heavy, sustained, privacy-sensitive use.

What It Does Not Do

Be clear about the ceiling. A post-trained 27B model running on a desktop is not Claude Opus 5 or GPT-5.6 Sol. On hard multi-step reasoning, long-horizon refactors, or anything requiring frontier capability, the local tier will underperform — which is precisely why the cloud escalation path exists.

The correct mental model is tiered, not replacement: cheap, private, unlimited local steps for the bulk of the loop; explicit, permissioned, billed cloud calls for the handful of steps that genuinely need frontier reasoning.

Also note what is absent at launch: no macOS, no AMD, no Windows, and no self-serve tier — you need a Perplexity Pro or Max subscription.

Why This Launch Matters Beyond Perplexity

Two things moved at once in late August 2026: Perplexity shipped a local agent harness on NVIDIA hardware (August 25), and Apple announced M6 and M5 Ultra with up to 512GB of unified memory (also August 25). Independently, both point the same direction — the desktop is being re-positioned as an inference device for agents, not just for chat.

The bottleneck for local agents was never really model quality; open-weight 27B-class models have been serviceable for a while. It was that nobody shipped the orchestration, sandboxing and tool plumbing in a form a normal user could install. Portable Computer is the first mainstream attempt to ship that layer as a product.

Should You Use It?

Yes, if: you already own a DGX Spark or a 24GB+ RTX card, run agents on code or documents you cannot send to a vendor, and have a Pro/Max subscription you are already paying for.

No, if: you are on a Mac, you need frontier reasoning quality, your usage is occasional, or you would be buying hardware specifically for this. At occasional volume the cloud wins on cost and wins outright on capability.