Perplexity Brings Local AI Agent to Windows RTX PCs

Perplexity has released Portable Computer, a local version of its AI agent that plans and executes multistep tasks directly on a user's own machine rather than routing them through cloud servers. Because the reasoning and execution loop runs on-device, sensitive files, credentials, and business data never have to leave the machine to get agentic help. Windows support launches today, joining existing support for Linux-based RTX PCs and NVIDIA DGX Spark systems; DGX Station support is expected to follow soon. The system requires an NVIDIA GeForce RTX or RTX PRO GPU with at least 24GB of VRAM, a threshold that immediately separates the hardware capable of running this class of agent from consumer cards that are not.

Under the hood, Portable Computer uses locally optimized models such as Qwen 3.8 27B, which removes the burden of manually configuring a local model stack. It ships with built-in connectors for Microsoft Office, Google Workspace, Slack, and GitHub, plus a browser and a sandboxed "SPACE" environment for running tasks. Perplexity highlights use cases like managing GitHub pull requests, running financial analysis across tax documents, and doing user-funnel analysis for startups, all without uploading files to an external service. No pricing was disclosed with the announcement.

Why the 24GB VRAM floor matters

A 24GB-or-more VRAM requirement is not a small detail; it is a filter. It rules out most mainstream consumer GPUs and points squarely at prosumer and workstation-class hardware, or dedicated systems like DGX Spark.

That is a deliberate tradeoff: running a capable agent loop, plus a 27B-parameter class model, plus browser and sandbox overhead, needs real memory headroom. The practical effect is that "local AI agent" stops being a marketing phrase anyone can claim and becomes a spec sheet, you either have the GPU memory to run it, or you don't. This sets a real hardware floor for what counts as a serious on-device agent, and it will likely push a wave of GPU upgrades among enthusiasts and small businesses that want the privacy benefit without paying cloud API bills per task.

A pragmatic middle path: local-first, cloud on request

The hybrid model, running locally by default but escalating to cloud-scale reasoning only with explicit user permission, is a middle path between two extremes that have each struggled to satisfy real users: pure-cloud agents that are powerful but require sending every document off-device, and pure-local agents that protect privacy but hit a capability ceiling on the hardest tasks.

By keeping escalation opt-in, Perplexity lets a user handle the bulk of routine document analysis, workflow automation, and data extraction entirely on their own machine, only reaching for more powerful cloud models when a task genuinely needs it. That also reduces dependency on cloud service credits and per-call API costs for everyday work, since most tasks never leave the device in the first place.

Enterprise data residency and privacy-sensitive workflows

For workflows in law, tax, and finance, where documents often cannot leave a jurisdiction or a firm's own infrastructure for compliance reasons, a local-first agent changes what is possible.

The financial-analysis-across-tax-documents example Perplexity cites is a direct illustration: an agent that can read, cross-reference, and summarize sensitive filings without transmitting them anywhere is a materially different risk profile than a cloud-only tool, even one with strong data-handling promises. The same logic applies to managing GitHub pull requests or analyzing a startup's user funnel, tasks that touch proprietary code or internal metrics that many teams would rather not route through a third-party server by default.

Competitive context

Running a full agent stack on consumer and prosumer GPUs, rather than requiring a dedicated cloud subscription, puts Perplexity in a small but growing group of vendors betting that on-device inference has crossed a usability threshold. The requirement for RTX or RTX PRO GPUs with 24GB-plus VRAM, or NVIDIA DGX Spark and DGX Station hardware, ties the offering tightly to NVIDIA's hardware roadmap rather than treating the GPU as an interchangeable commodity.

As more vendors chase the same local-agent niche, the real competition may end up being fought less on model quality and more on how gracefully each product handles the local-to-cloud handoff, and how low the VRAM floor can go without gutting capability. Today's launch, with no pricing yet disclosed, does not settle that question, but it does mark one of the more concrete steps yet toward agentic AI that treats local execution as the default rather than the exception.