Perplexity’s Portable Computer, a local version of its agentic AI platform, is now available on Windows PCs with NVIDIA GeForce RTX and RTX PRO Workstation GPUs, the companies announced on September 14. The release expands a platform that previously ran only on NVIDIA’s DGX Spark desktop system and RTX machines running Linux.
Portable Computer is a local-first version of Perplexity Computer, the company’s agent that plans and carries out multistep tasks. Running on device, it uses local models to analyze data, pull together information across files and handle recurring work. Sensitive information stays on the machine, and locally completed work does not consume Perplexity Computer credits. Users can still escalate specific tasks to cloud models for heavier research and reasoning.
Why local matters
The pitch targets a specific pain point: the most valuable work people want agents to do, from private codebases to confidential business documents, is exactly the work they are least willing to send to a cloud service. Perplexity introduced Portable Computer on August 25 alongside NVIDIA’s DGX Spark, a desktop-sized AI computer built around the Grace Blackwell GB10 platform with a 20-core Arm CPU and 128 GB of unified memory. The first release ran on Linux only, with Windows support promised for later.
That promise landed on schedule. Bringing the product to standard Windows RTX PCs matters because it drops the hardware requirement from a specialized desktop system to machines millions of people already own. Perplexity’s own framing in the launch post was that running AI on personal machines is going to be a much bigger part of how work gets done as local models get stronger and chips get faster. The company positions every chip cycle and model release as pushing that shift forward.
For Perplexity, the release is also a way to differentiate its agent from cloud-only rivals. Anthropic, OpenAI and Google all run their agents server-side, which means customer files leave the building by default. A local-first option flips that default, and it aligns with the caution enterprises have shown since a series of AI data-handling incidents made procurement teams ask where their documents actually go. Law firms and healthcare organizations in particular have been slow to adopt agentic tools for exactly this reason, and a vendor that can promise on-device execution removes the largest single objection.
What runs where
The on-device stack is more complete than a chatbot wrapper. Perplexity says the orchestrator, planner, tool router, scheduler, durable task queue and local search index all run locally. On DGX Spark the platform uses Qwen 3.8 27B or PPLX 27B, a post-trained version of the Qwen model, with NVIDIA’s Nemotron 3.5 Lightning 30B open model planned for the model picker. Dictation runs locally too, on the Nemotron 3.5 ASR model, so a user can talk through a private codebase or dictate over client documents without the audio ever touching the cloud.
When a task needs the outside world, the agent can call Perplexity’s cloud search or deep research and return cited findings, while everything else stays on the machine. The company says security matches the cloud version of Perplexity Computer. Installation is a one-click setup through the Perplexity app.
Part of a bigger local-AI push
NVIDIA has been building out a local AI ecosystem around its RTX and DGX lines, and the blog post announcing the Windows release sits alongside a string of recent launches: Z.ai’s GLM 5.3 Flash optimized for DGX Station and dual DGX Spark systems, DeepSeek-v4.1 Flash with reduced key-value cache memory demands for agent workloads, Qwen’s open-weight Qwen3.8-Flash-Next with an early Qwen4 preview that runs on a single DGX Spark, and GLM 5.3, Z.ai’s 744-billion-parameter flagship tuned for agent sessions that run for hours on a four-Spark cluster.
The timing also fits the hardware cycle. ASUS and MSI burned through first production runs of RTX Spark laptops before retail launch this month and ordered 100,000 more, a sign that local AI capability is selling machines rather than just filling spec sheets. Nvidia’s RTX branding has quietly shifted from gaming-first to AI-first in its marketing, and a marquee software partner like Perplexity anchoring the Windows experience helps that story.
The tradeoffs
Local models remain a tier below frontier cloud models on complex reasoning, and a 27B-class model on a PC will not match what GPT-6 Astra or Claude’s flagship can do on hard analysis. That is why the hybrid design keeps cloud escalation available. The bet is that for the daily bulk of agent work, file analysis, recurring tasks, document synthesis, on-device models are good enough, and the privacy and cost advantages outweigh the capability gap.
There is also a hardware floor. Even with RTX GPUs, memory and VRAM constrain what a Windows PC can run. NVIDIA’s unified-memory DGX design exists precisely because ordinary PCs cannot hold large models comfortably. Users with older 8 GB cards will not get the full experience, and Perplexity has not published minimum specs for the Windows build beyond compatible RTX hardware.
Portable Computer on Windows is available to Perplexity Pro and Max subscribers on compatible GeForce RTX and RTX PRO machines. The enterprise angle is the quieter part of the story: firms handling regulated or confidential data want the option to keep processing local, and this consumer release builds the same muscle. If the hybrid local-plus-cloud pattern holds, the next frontier in the AI race may be decided less by model quality than by how much work never leaves the user’s desk.