Perplexity Portable Computer Runs Local AI on Windows

The interesting thing about Perplexity's Portable Computer is not that it runs a language model locally — plenty of tools do that. It is that the whole agent runs locally. The model, the agent harness, the orchestrator and the scheduler all sit on the machine, which means a multi-step task can plan, execute and repeat itself without a round trip to anyone's data centre. On September 14, 2026, that stack arrived on Windows.

• Portable Computer for Windows launched September 14, 2026, running on NVIDIA GeForce RTX and RTX PRO GPUs with 24GB or more of VRAM
• Ships with Qwen 3.8 27B, post-trained for Perplexity Computer and optimised for RTX, alongside Perplexity's in-house PPLX 27B
• NVIDIA Nemotron 3.5 Lightning, a 30-billion-parameter model, is listed as incoming, with DGX Station support coming
• Connectors cover Microsoft Outlook, OneDrive and Word, plus Google Drive, Gmail, Slack and GitHub, with scheduled recurring tasks and local MCP servers

What Does the 24GB VRAM Requirement Actually Buy You?

That threshold is the story in one number. A 27-billion-parameter model at a sensible quantisation needs room for weights plus a working context, and agentic runs chew through context fast — tool outputs, file contents, intermediate plans. Twenty-four gigabytes is roughly where a model of this class stops thrashing and starts behaving like a dependable worker rather than a demo.

It also draws a hard line around who can run this today. A GeForce RTX card with 24GB or more, or an RTX PRO workstation card, is enthusiast-and-above territory. NVIDIA says DGX Station support is coming, which extends the ceiling rather than the floor. For anyone sizing up hardware for this kind of work, our best mini PC for local LLMs guide covers the memory-bandwidth trade-offs that decide whether a machine feels fast or merely capable.

Which Models Run On-Device?

The default local model is Qwen 3.8 27B, which Perplexity post-trained specifically to drive Perplexity Computer and optimised for RTX GPUs. Sitting beside it is PPLX 27B, the variant Perplexity trained in-house. NVIDIA Nemotron 3.5 Lightning, at 30 billion parameters, is listed as arriving.

That is a meaningful pattern: a frontier-lab-adjacent product shipping with an open-weight Chinese model as its default local engine, tuned for the specific agent loop it has to drive. The open-weight tier has quietly become the substrate for on-device agents, because you cannot ship a closed API model inside someone else's GPU. Users can still reach for cloud models when a task genuinely needs deeper reasoning — the local path is a default, not a cage.

Why Local Agents Are a Privacy Argument, Not Just a Speed One

The pitch NVIDIA leads with is that sensitive data stays on the device. That is not marketing filler for the workloads listed: reviewing pull requests, consolidating information across files, analysing financial records, spotting gaps in documentation. Each of those involves material that many organisations simply will not send to a third-party endpoint, regardless of the provider's retention policy.

The local MCP server support matters here too. Model Context Protocol connectors let the agent reach desktop applications directly — Outlook, OneDrive, Word, Google Drive, Gmail, Slack, GitHub — with the connection terminating on the same machine. Add scheduled recurring tasks and you have something closer to a resident assistant than a chat window: a process that wakes up, does the work, and leaves the results where you can find them.

How to Get It

Access runs through the existing Perplexity app for Windows, distributed via the Microsoft Store, and is available to Pro and Max subscribers on both individual and enterprise plans. The hardware gate is the binding constraint, not the subscription tier.

This is the same direction of travel we tracked when DeepSeek V4.1 Flash shipped open weights with a one-million-token context: capable models are getting small enough and cheap enough that the default deployment target is shifting from someone else's rack to the desk in front of you. Our full AI coverage follows that shift as it lands.

Sources: NVIDIA Blog — September 14, 2026; Tom's Hardware — September 2026.