Most AI agents today have a simple problem: your work lives on someone else's servers. This week, Perplexity and NVIDIA took a real swing at changing that. Perplexity Portable Computer, the company's fully local AI agent, is now available on Windows PCs — no Linux installation, no separate DGX box, just your own NVIDIA RTX GPU doing the heavy lifting.
01.What is Portable Computer, anyway?
Portable Computer is the local version of Perplexity Computer, the company's agent for multi-step tasks — the thing that plans a job, runs subtasks through connectors and tools, and hands you finished work instead of a chat reply. The cloud version runs on Perplexity's infrastructure and consumes Computer credits. The portable version runs the whole stack — the model, the agent harness, the orchestrator and the scheduler — on your own machine. It first launched on Linux and NVIDIA's DGX Spark desktop in August; until now, Windows users were left out.
02.What's new in the Windows release
The release landed on September 14 through the existing Perplexity app for Windows, and it brings a few genuinely useful additions. There are scheduled recurring tasks, so the agent can keep working on a rhythm without you babysitting it, and local MCP servers for desktop apps. Connectors cover Microsoft Outlook, OneDrive and Word, Google Drive and Gmail, Slack and GitHub. Code and tool execution happen inside isolated sandboxes, including Perplexity's own SPACE sandbox — the same containment the cloud product uses.
The default local model is Qwen 3.8 27B, post-trained by Perplexity to work with the Computer harness and optimised for RTX GPUs. Perplexity also ships PPLX 27B, its own post-trained variant of the same base, and NVIDIA's Nemotron 3.5 Lightning 30B is listed as coming soon. On Perplexity's internal Local Knowledge Work Bench, PPLX 27B scores 85.4% against Qwen 3.8 27B's 82.6% — a real gain from post-training, even if it's a benchmark the company built and grades itself.
03.The hardware catch: 24GB of VRAM
Here's the part that decides whether this matters to you. The host system needs a GeForce RTX or RTX PRO GPU with at least 24GB of VRAM, plus a Perplexity Pro or Max subscription. On the consumer side, that means an RTX 3090 or 3090 Ti (24GB), an RTX 4090 (24GB), or an RTX 5090 (32GB). On the workstation side, the RTX PRO 4000 through 6000 cards qualify. In India, a used 3090 still typically costs well over ₹70,000, and a new 4090 or 5090 goes far beyond that — so this is firmly an enthusiast-and-pro release, not a mainstream one. Support for NVIDIA's DGX Station is coming, and RTX Spark Windows PCs from Lenovo and Acer are expected in October.
04.Private by design — mostly
The privacy pitch is the strongest part. Sensitive files stay on your device, and when a task genuinely needs stronger reasoning or live web information, the agent can reach Perplexity Search or frontier cloud models — but it asks for your permission before anything leaves the machine. Perplexity CEO Aravind Srinivas called it "unmetered local intelligence on every Windows PC running on NVIDIA hardware and Perplexity harness." That's a notable line, because the browser and OS world is moving the same direction — Mozilla just put Mistral's open models inside Firefox for similar privacy reasons.
05.The economics are the real pitch
Work the agent finishes locally consumes no Perplexity Computer credits. Think about what that means in practice: an agent that spends an hour grinding through a document review or an iterative analysis racks up a real bill against a metered cloud API — and close to nothing against a GPU you already own. Perplexity's own measured compromise is telling: handing a difficult step up to a frontier model like Claude Opus 5 recovered roughly 60% of the gap to frontier performance at about two-thirds the cost of running the whole task in the cloud. Local first, cloud when needed.
If you're new to this whole agentic-AI thing, by the way, our GPT-6 Astra — Zero to Hero Guide has a full chapter on how agents and computer use actually work under the hood — the same concepts Portable Computer is built on — plus a cost-optimization section that'll save you far more than the guide's price.
06.Should you care in India?
Honest answer: not yet, unless you already own qualifying hardware or work with genuinely sensitive data. The 24GB VRAM floor rules out nearly every consumer PC sold in India, including every mainstream laptop. But the direction matters. Google shipped Gemini 3.8 Live earlier this week for cloud voice agents; Perplexity is betting the opposite way — that the agent runtime moves down the stack and onto the device. For CA firms handling client financials, clinics, or anyone who can't put data on a US server, local agents like this will eventually stop being a luxury.
Key Takeaways
- Perplexity Portable Computer is now on Windows via the Perplexity app — previously Linux/DGX Spark only
- Requires an RTX or RTX PRO GPU with 24GB+ VRAM and a Pro or Max subscription
- Runs Qwen 3.8 27B / PPLX 27B locally; local work burns zero Computer credits
- Nothing goes to the cloud without explicit permission; code runs in isolated sandboxes
- New in the Windows build: scheduled recurring tasks and local MCP servers
Frequently Asked Questions
What is Perplexity's Portable Computer?
Perplexity has brought Portable Computer — its fully local AI agent — to Windows PCs with NVIDIA GeForce RTX or RTX PRO GPUs. The entire agent stack runs on your machine, so local work burns zero Perplexity credits.
What are the requirements for Portable Computer on Windows?
You need a GPU with at least 24GB of VRAM, plus a Perplexity Pro or Max subscription. Nothing leaves your PC without explicit permission — the agent asks before escalating a task to the cloud.
What This Means for You
If you're a developer or founder in India, file this one away rather than acting on it today. The subscription-plus-local-GPU combination doesn't make sense at Indian hardware prices yet — but the pattern does. Agents that work offline, keep data on your machine, and stop metering every token are coming down in hardware requirements every quarter. The folks who win when that hits mainstream won't be the ones with the best GPUs; they'll be the ones who already know how to direct agents properly. That's a skill you can build now, and our GPT-6 Astra — Zero to Hero Guide is the cheapest way we know to do it.