Why use Portable Computer
- Keep sensitive work on the device. Files, queries, and agent activity stay on your machine unless you approve sending something to the cloud.
- Run long or recurring work without spending credits. Work handled by the local model does not consume Computer credits, which makes long-running and scheduled workflows practical.
- Schedule work on your own hardware. Recurring jobs such as a daily pull request review or a weekly summary of local files run on the device on a schedule you set.
- Bring in the cloud only when you need it. When a task needs current information or advanced reasoning, Portable Computer asks permission before routing that step.
- Use tools running on your machine. Local MCP servers let Portable Computer drive tools installed on your device, which cloud Computer does not do.
Portable Computer and Personal Computer
These are two different features, and both can be active at the same time.
Table 1 — How Portable Computer differs from Personal Computer.
On Windows, Personal Computer is the app you install; Portable Computer is the
local inference you then turn on inside it. You can use Personal Computer without
Portable Computer, and local file access works either way.
What runs on your device
Portable Computer is not a thin client. The following run locally:- The local model that acts as orchestrator
- The planner and tool router
- The scheduler and durable task queue
- The local search index over your files and code
- The sandbox that isolates code and tool execution
Local models
Table 2 — Local models and the hardware they run on.
Only one local model runs at a time.
What you need
Table 3 — Requirements for running Portable Computer.
The 24 GB VRAM requirement applies to the RTX GPU option. DGX Spark meets it on
its own. If your GPU reports less, on-device models and their download controls
are hidden and the Local Inference page shows a notice instead.
If an organization manages your account, an administrator may need to allow
local inference before you can use Portable Computer.
Disk space
Table 4 — Approximate download size by platform and model.
Leave extra room beyond the download size. Files are unpacked after downloading,
and model updates need the new model’s size plus about 5 GB of headroom.
Set up Portable Computer
On Windows there is no separate installer, because local inference lives inside the app you already have. Downloading a model is what sets it up, and selecting that model in the composer is what switches a task to run locally.To set up Portable Computer on Windows
1
Install the Perplexity app for Windows from the Microsoft Store or the
Perplexity download page, then sign in. For detailed steps, see Get started
with the desktop apps.
2
Open the Menu at the top left, then select Settings.
3
Select Local Inference.
4
Under On-device models, select the download button next to PPLX 27B.
The app downloads the model and installs the local inference engine.
5
When setup finishes, open the model menu in the composer and choose PPLX
27B under Local.
Before setup, the Local section of the model menu shows Set up instead of
a model.
On Windows, local file access is a separate setting from local inference. To let
Portable Computer work with files on your device, enable Local file system
access and turn on Desktop tools in Settings → Permissions. See Get
started with the desktop apps for the full
steps.
To set up Portable Computer on DGX Spark or Linux
1
Add the Perplexity package repository and install the package:
2
Open Perplexity and sign in.
3
Open the Menu, select Settings, then select Local Inference.
4
Under On-device models, select the download button next to PPLX 27B
or Qwen 27B.
5
When setup finishes, open the model menu in the composer and choose your
model under Local.
Ubuntu and DGX OS are the officially supported Linux distributions.
To confirm Portable Computer is running locally
- Check that a local model is selected under Local in the model menu. Tasks run on the device while it is.
- Watch for the approval prompt. If a step needs the cloud, Portable Computer asks before sending anything from your device.
To switch between local and cloud
- To run a task locally, select your local model under Local in the model menu before you start.
- To run a task in the cloud, select a cloud model instead.
- To allow a single cloud step inside a local task, approve the prompt when Portable Computer asks.
Escalating to the cloud
The local model only knows what is on your machine. When a task needs more, Portable Computer can escalate a single step:- Perplexity Search, wide research, and deep research return cited findings from the web into the same task.
- Frontier models handle advanced reasoning, with a choice of 15+ models.
- Connected apps and browser use run through the cloud when a step requires them.
Portable Computer asks for your permission before sending anything from your
device to the cloud. A step you approve runs in the cloud and uses Computer
credits at the standard rate.
Working with apps and local tools
Portable Computer works with connected apps through the local orchestrator, including Microsoft Outlook, OneDrive, and Word, plus Google Drive, Gmail, Slack, and GitHub. It also supports MCP servers running on your own machine, so it can drive local tools. These run on your device and are used only in Local mode. Cloud Computer uses separately managed cloud connectors and does not use servers configured under Settings → MCP servers.To add a local MCP server
1
Set up a local model and select it under Local in the model menu.
2
Open the Menu, select Settings, then select MCP servers.
3
Select Add server.
4
Choose Local (stdio).
5
Enter a Name and the absolute path to the Command. Use Browse to
locate it.
6
Add Arguments one per line, and a Working directory or Environment
variables if the server needs them.
7
Select Test to check the connection, then select Add.
The Command must be an absolute path. Portable Computer does not search your
PATH.Scheduled tasks
You can schedule recurring work to run on the device. A scheduled task uses the local scheduler and durable task queue, so it keeps running on your hardware rather than in a cloud sandbox.Controls and limits
- Local work does not consume credits. Steps you approve for cloud escalation are billed like other cloud Computer work.
- Nothing leaves the device without approval. You are prompted each time a task needs to send content to a cloud service.
- Code and tools run in a sandbox. Execution is isolated, with controlled access to your files and connected apps.
- One local model at a time. Switching models means loading a different one, not running both.
- Brain is a cloud feature. Brain and saved memory are not available in local mode.
- The local model is smaller than a frontier model. For advanced reasoning, escalate the step.
If setup does not complete
Table 5 — Setup symptoms, with the most likely cause and fix.