Perplexity Portable Computer Review: Local AI With a 24GB Catch

This Perplexity Portable Computer review examines an unusually complete attempt to move an AI agent from the cloud onto hardware you control. Portable Computer combines a local model, agent harness, orchestrator, sandbox, connectors, search, and optional cloud escalation. Its September 2026 Windows release makes that idea more accessible, but the 24GB VRAM minimum means it is still a specialist product rather than an ordinary laptop assistant.

Perplexity Portable Computer Review: Quick Verdict

Portable Computer is most interesting for developers, researchers, analysts, and small teams that already own a supported NVIDIA system and regularly work with private files or large local datasets. It packages more of the agent stack than a do-it-yourself local model setup, supports long-running and scheduled work, and does not charge cloud credits for tasks completed locally.

The catch is substantial. Windows and Linux PCs need a supported NVIDIA RTX or RTX PRO GPU with at least 24GB of VRAM, or users can run it on NVIDIA DGX Spark. A Perplexity Pro or Max subscription is also required. Local execution can reduce data exposure and per-token fees, but it does not make the system free, fully offline, or automatically safe. Connectors, web research, and approved cloud escalation can still cross the device boundary.

This is an evidence-based review, not a hands-on benchmark. Product behavior and benchmark results below are attributed to Perplexity or NVIDIA where appropriate.

What Is Perplexity Portable Computer?

Portable Computer is a local-first version of Perplexity Computer, the company's agent for multistep knowledge work. Instead of sending every task to a hosted model, it runs the agent harness, orchestration, model inference, and sandbox on a supported Windows or Linux machine. Perplexity says the agent can analyze local files, execute code and tools in isolated environments, use connected apps, schedule recurring work, and produce documents or research outputs.

Windows support arrived on September 14, following the August Linux and DGX Spark launch. On Windows, users begin in the Perplexity app and enable PPLX 27B for local inference in Settings. Linux users install Perplexity's package and download a supported local model.

Hardware and Software Requirements

The official product page lists three practical routes: NVIDIA DGX Spark, a supported Linux PC with an NVIDIA RTX GPU, or a supported Windows PC with NVIDIA GeForce RTX or RTX PRO graphics. The RTX paths require at least 24GB of VRAM. That excludes most mainstream gaming laptops and many otherwise fast desktop cards.

Windows and Linux RTX systems currently use PPLX 27B. DGX Spark can run PPLX 27B or Qwen 3.8 27B, while NVIDIA Nemotron 3.5 Lightning is listed as coming soon. Only one local model runs at a time. Support should be checked against the live compatibility page before buying hardware, because model availability and GPU eligibility can change.

How the Local-First Workflow Works

Local Models, Harness, and Sandbox

PPLX 27B is Perplexity's post-trained model based on Qwen 3.8 27B. The surrounding Computer harness manages planning, tools, context, and task execution. Perplexity says code and tool use run in isolated sandbox environments with controlled access to files and connected apps.

Search, Connectors, and Cloud Escalation

When a task needs frontier reasoning, the orchestrator can propose sending a step to one of more than 15 cloud models. Perplexity and NVIDIA say the user is asked before that handoff. Treat the approval screen as a privacy boundary: review exactly what information will leave the machine, which service will receive it, and whether the task can be completed locally instead.

Privacy and Security: Local Does Not Mean Isolated

Portable Computer's strongest privacy benefit is data minimization. Work completed locally does not need to send every prompt, file, and intermediate token to a remote inference provider. This can be valuable for source code, contracts, financial records, or unpublished research. It may also make auditing easier because the compute and workspace remain on hardware the user controls.

However, “local-first” is more accurate than “offline.” Search requests, connectors, software updates, and approved cloud escalation create network paths. Posting a summary to Slack or reading Gmail necessarily involves those services. A sandbox reduces exposure but is not proof that an agent cannot make a harmful tool call, follow a malicious instruction in a document, or send information to the wrong destination.

Use a dedicated workspace, grant only the connectors the task needs, separate read and write permissions, and require confirmation for external messages or file changes. SD's AI agent security guide explains why prompt injection, credential scope, logging, and human approval still matter for local agents.

Performance Evidence and Its Limits

Perplexity reports that PPLX 27B scored 85.4% on its Local Knowledge Work Bench, a set of 53 tasks across research, data, document creation, engineering, compliance, visualization, and project work. The company says its Computer harness with the base Qwen 3.8 27B scored 82.6%, compared with 77.6% for Pi and 74.0% for Hermes in the same DGX Spark evaluation.

These results are useful engineering evidence, not an independent product verdict. Perplexity created the benchmark, post-trained the model, designed the harness, and reported the scores. It says each task was run three times and that a held-out subset was used for final evaluation, but the benchmark and technical report were not yet fully open at publication time. Results on a user's files, connectors, Windows GPU, or long-running workflow may differ.

Pricing and the Real Cost of Running Locally

Portable Computer has no separate software price on the current product page, but access requires a qualifying Perplexity Pro or Max subscription. Perplexity lists Max at $200 per month or $2,000 per year; regional prices and plan entitlements should be checked before purchase. Enterprise plans have separate seat pricing and data terms.

Local tasks do not consume Perplexity Computer cloud credits, according to the company. That can make high-volume document or repository work more predictable. Cloud model escalation and other hosted features may still use allowances or credits, so “no token cost” applies only to work that stays local.

Perplexity Portable Computer Pros and Cons

Pros

• Local inference can keep sensitive files and most intermediate work on the device.
• The product bundles a model, agent harness, sandbox, connectors, scheduling, and research tools.
• Locally completed work does not consume cloud credits.
• Windows and Linux support broaden access beyond DGX Spark.
• Permissioned cloud escalation offers a practical path when the local model is not enough.

Cons

• The 24GB VRAM floor excludes most consumer PCs and laptops.
• A paid Perplexity subscription is still required.
• Connectors and cloud steps complicate the privacy boundary.
• The strongest performance evidence comes from Perplexity's own benchmark.
• Hardware, power, setup, and maintenance can outweigh saved API costs for light users.

Who Should Use Portable Computer?

Portable Computer fits people who already have suitable NVIDIA hardware, need an agent to work across private local files, and prefer a packaged experience over assembling a model server and agent framework. Good pilots include repository analysis, document reconciliation, internal research, and scheduled reporting where outputs remain reviewable.

It is a weaker fit for casual users, Mac-only teams, or anyone buying an expensive GPU for occasional summaries. Users who mainly want AI-assisted browsing may prefer the product covered in SD's Perplexity Comet review. Technical operators who value open configuration over a managed stack should compare SD's OpenClaw review, while hardware buyers can use the Surface RTX Spark Dev Box overview for broader local-agent workstation context.

How to Evaluate It Before Committing

Choose one representative task and prepare a small evaluation set with normal inputs, missing data, conflicting files, and a malicious or misleading instruction. Run it locally several times. Record completion rate, time, corrections, GPU memory use, power draw, and any proposed cloud escalation.

Then enable one connector with the narrowest practical permission. Check whether the agent chooses the right source and destination, shows a meaningful approval before external action, and leaves an understandable audit trail. Compare the result with a cloud agent and a simple local-model setup. The winning option is the one with the lowest total cost per accepted task, not the most impressive architecture diagram.

Conclusion

This Perplexity Portable Computer review finds a credible local-first agent for people who need private file access, long-running workflows, and a managed stack on hardware they control. Windows support is a meaningful expansion, and bundling the local model, harness, sandbox, connectors, and cloud handoff removes substantial setup work.

The 24GB VRAM requirement keeps it niche, while connector access and cloud escalation mean privacy still depends on configuration and review. Treat Perplexity's benchmark as vendor evidence, start with a narrow reversible workflow, and compare total hardware and subscription cost with the cloud alternative. Portable Computer is most convincing when the required GPU already exists and sensitive, high-volume work can genuinely remain local.