AnswersPublished21 sources
Perplexity Portable Computer Runs AI Agents Locally—With Cloud Help Only When You Approve
Launched on August 25, 2026, Perplexity Portable Computer runs its models, orchestrator, tools, and agent harness locally, so private files can stay on the device and local work does not consume Perplexity billing cre... When a task needs a frontier cloud model, Portable Computer is designed to ask for permission be...
Published byEdited with GPT-5.6 LunaImages generated with GPT Image 1.5
Research answer
Key takeaways
- Launched on August 25, 2026, Perplexity Portable Computer runs its models, orchestrator, tools, and agent harness locally, so private files can stay on the device and local work does not consume Perplexity billing cre...
- When a task needs a frontier cloud model, Portable Computer is designed to ask for permission before sending that individual step off device rather than routing the entire workflow to the cloud automatically.
- At launch, the available local models are Qwen 3.8 27B and PPLX 27B; Perplexity reports strong internal benchmark results, but those figures have not been independently validated in the supplied sources.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is Perplexity’s Portable Computer, the local AI agent launched in partnership with Nvidia in August 2026, how does it preserve privacy. Article summary: Perplexity’s Portable Computer is a local-first version of its Computer agent platform, launched with Nvidia on August 25, 2026. It packages an AI agent runtime for Nvidia DGX Spark and qualifying Linux RTX systems so se. Topic tags: general, general web, user generated. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, charts with fak
Perplexity Portable Computer is a local-first version of the company’s Computer agent platform, launched in partnership with NVIDIA on August 25, 2026. Instead of sending an entire multi-step workflow to remote servers, it packages the agent runtime—including models, planning, tools, connectors, and sandboxing—for supported local hardware. 3
5
The result is a hybrid design: everyday agent work can happen on the user’s machine, while more demanding reasoning can be sent to a frontier cloud model only after the user approves the escalation. 3
6
What makes Portable Computer different from a local chatbot?
Portable Computer is intended to run more than just inference locally. Its package includes the orchestrator or planner, subagent models, agent harness, inference engine, tools, app connectors, and a security sandbox. That means the operational workflow—not merely a standalone language model—can remain on the device. 3
5
This architecture extends Perplexity’s broader Computer concept, which is designed to execute multi-step work across files, applications, connected services, and the web. The local version is aimed at keeping sensitive inputs and routine processing closer to the user’s hardware while preserving the same agent-style workflow. 3
25
How Portable Computer preserves privacy
The default path is local execution. Perplexity says tasks begin on the user’s machine, where the local model can work with files and tools without automatically sending the underlying work to the cloud. Locally completed work also does not count toward Perplexity token limits or billing credits. 1
3
Potential local workflows include:
- Reviewing private folders containing tax forms or brokerage records
- Analyzing CSV files and other documents
- Extracting information and producing reports
- Creating documents and other outputs
- Working with connected services such as Slack, Google Drive, Gmail, and GitHub
1
3
8
Portable Computer also uses an OS-level security sandbox for the agent harness, according to Perplexity’s launch coverage. If the required sandbox is unavailable, the harness disables tool execution instead of running without that protection. 3
5
That is a meaningful distinction from a cloud-first agent: local processing can reduce the amount of sensitive material that needs to leave the machine. It is not, however, a guarantee that no data will ever reach a server, because the product can request cloud assistance for subtasks that exceed the local model’s capabilities.
How hybrid cloud escalation works
Portable Computer is local-first rather than local-only in every circumstance. If a subtask requires a more capable frontier model, the system is designed to ask before sending that individual step to the cloud. The cloud response then returns to the local workflow. 3
6
This user-gated approach gives users a choice between local privacy and cloud-based reasoning power. The supplied reporting supports the approval-based escalation model, but it does not independently confirm every more specific implementation detail described in the original request, including PII-flagging behavior or a particular text-only approval process. Those details should therefore be treated as unverified here.
Which local models does it support?
At launch, Portable Computer supports two local model options:
Perplexity said support for NVIDIA Nemotron 3.5 Lightning was planned next. 3
8
The relatively focused model selection is part of the trade-off for running the system locally. Users get a packaged agent experience rather than access to the full range of frontier models available through a cloud platform, but they can keep more routine work on hardware they control.
What performance does Perplexity report?
Perplexity reports results from its internal 53-task Local Knowledge Work Bench showing:
- 82.6% for Qwen 3.8 27B with Computer
- 85.4% for PPLX 27B
- 77.6% for Pi using the same model
- 74.0% for Hermes using the same model
The company also reports 66.7% on BrowseComp, compared with 50.2% for Pi and 43.9% for Hermes. In the same comparison, Perplexity reports 51% less wall-clock time and 70% fewer tokens than Pi. For multimodal document understanding, it reports 65.1%, compared with 34.6% for Pi and 13.9% for Hermes. 3
These figures are vendor-run results from Perplexity’s reported evaluations, not independent benchmark tests. They are useful for understanding the company’s performance claims, but they should not be read as conclusive proof that Portable Computer will outperform other local-agent systems in every workload.
Hardware and operating-system requirements
At launch, Portable Computer supports:
- NVIDIA DGX Spark, and
- Linux machines with an NVIDIA RTX GPU and at least 24 GB of VRAM, roughly the class of an RTX 3090 or newer.
5
15
Windows support was expected in September 2026. 2
3
The hardware requirement is the clearest limitation of the launch. Local agent workflows need substantially more compute and memory than ordinary cloud-connected chat, so Portable Computer is initially aimed at users with high-end NVIDIA hardware rather than the typical laptop owner.
Which Perplexity plans include it?
Portable Computer is included at no additional cost for subscribers on these tiers:
- Pro
- Max
- Enterprise Pro
- Enterprise Max
“No additional cost” refers to the product entitlement; users still need compatible hardware to run the local software.
How it extends Perplexity Computer
Perplexity introduced Computer as a cloud-oriented agent capable of coordinating models, browsing, using tools, and completing extended workflows. 25
28 Portable Computer brings that same general agent direction to local NVIDIA hardware rather than treating on-device inference as a separate chatbot.
It also follows the July 2026 Windows release of Personal Computer, which brought Perplexity’s agent platform to local files, Microsoft 365, and the web on Windows. 18
The strategic progression is straightforward:
- Computer established the cloud-based multi-step agent.
- Personal Computer brought that agent closer to users’ local files and desktop applications.
- Portable Computer moves the central agent runtime itself onto supported local hardware, with optional cloud assistance when local models are not enough.
18
The trade-off: privacy and control versus convenience
Portable Computer’s strongest selling point is not simply that it runs a model on a GPU. It bundles the model, orchestrator, harness, inference engine, tools, connectors, and sandbox into a more complete local-agent product. That reduces the setup burden associated with assembling a local stack from separate components. 3
5
Its compromise is equally clear: users need powerful NVIDIA hardware, the initial operating-system support is limited, and local models may not match the reasoning range of the newest cloud systems. The hybrid approval flow is intended to offer a middle ground—keep sensitive and routine work local, then use cloud intelligence for selected steps when the user agrees.
For people handling private documents or business data, that privacy-first default may matter more than having unrestricted access to the largest cloud models. For everyone else, the hardware gate and limited launch availability may make a conventional cloud agent more convenient until support expands.
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Perplexity Portable Computer Runs AI Agents Locally—With Cloud Help Only When You Approve"?
Launched on August 25, 2026, Perplexity Portable Computer runs its models, orchestrator, tools, and agent harness locally, so private files can stay on the device and local work does not consume Perplexity billing cre...
What are the key points to validate first?
Launched on August 25, 2026, Perplexity Portable Computer runs its models, orchestrator, tools, and agent harness locally, so private files can stay on the device and local work does not consume Perplexity billing cre... When a task needs a frontier cloud model, Portable Computer is designed to ask for permission before sending that individual step off device rather than routing the entire workflow to the cloud automatically.
What should I do next in practice?
At launch, the available local models are Qwen 3.8 27B and PPLX 27B; Perplexity reports strong internal benchmark results, but those figures have not been independently validated in the supplied sources.
Sources
tomsguide.comPerplexity's new local-first AI runs on your PC — and asks before using the cloud