Perplexity and NVIDIA on August 25 jointly launched "Portable Computer," a local-first AI agent that runs entirely on your own hardware instead of in the cloud. Packaged as a free add-on to Perplexity's paid subscriptions, it bundles everything needed to run agentic AI locally: the model, the server, and the harness.
The pitch is simplicity. Until now, running local AI meant juggling drivers, model weights, an inference server, and a tool harness, a steep climb for anyone without a software development background. Portable Computer collapses all of that into a single install that downloads what you need in one shot, so a novice can buy a compatible machine and start using an agent the same day.
The launch lands at a moment when local AI is suddenly practical for more people. Cheap unified-memory machines like the DGX Spark are bringing workloads to the desktop that were once cloud-only, yet the software layer has lagged behind the hardware. Perplexity and NVIDIA are betting that a turnkey agent stack, rather than another DIY toolchain, is what finally unlocks adoption.
What Is Portable Computer?
Think of it as Perplexity's answer to setting up local AI for ordinary users. The package includes the models, a serving stack, and Perplexity's agent harness, which connects to the tools you already use. At launch the harness supports Gmail, Google Drive, Slack, and GitHub through what Perplexity calls app connectors, which AI-savvy users will recognize as MCP servers.
- Connect to Gmail, Google Drive, Slack, and GitHub
- Run shell commands under human supervision, read PDF and text files, and take dictation through the Nemotron 3.5 speech model
- Search the web and analyze content, then escalate to cloud frontier models only with your permission
Every task begins on your machine, so the local portion runs at no cost beyond your subscription.
What It Runs On and What It Does
The initial release targets NVIDIA's GB10 platform, meaning hardware like the DGX Spark or the Dell Pro Max with GB10. It also supports x86 PCs running Ubuntu with an NVIDIA GPU that has at least 24 GB of VRAM, which covers the RTX 3090, the RTX 4090, and the 32 GB RTX 5090. Windows support is expected in September, which should extend reach to NVIDIA's RTX Spark laptops.
- NVIDIA DGX Spark and Dell Pro Max with GB10
- x86 PCs on Ubuntu with at least 24 GB of NVIDIA VRAM, including RTX 3090, 4090, and 5090 cards
- Windows machines arriving in September
Launch support is limited to one model: Alibaba's Qwen3.8 27B, a dense 27-billion-parameter model released earlier this month. Perplexity says a Nemotron 3.5 Lightning model is coming soon, a 30-billion-parameter hybrid mixture-of-experts design that lights up only about 3 billion parameters per token, a far lighter load on a 24 GB card. Perplexity also post-trains its models for the specific tasks Portable Computer targets.
If the local model hits a step it cannot handle, it can, with your explicit permission, call out to Perplexity's hosted frontier models. Only that cloud portion consumes AI credits. Everything processed locally stays on your device and costs nothing extra.
The economics are what make it notable. Cloud agent usage is billed by the token, and heavy agentic workflows can burn through credits fast. By keeping the default path on-device, Portable Computer shifts that calculus: subscribers pay their normal plan fee, and only occasional cloud escalations add cost.
The Bottom Line
Portable Computer is rolling out to Perplexity Pro, Max, Enterprise Pro, and Enterprise Max subscribers, who receive it as part of their existing plan. For a field drowning in setup guides, the move stands out: an agent that works on hardware you already own, with zero token fees on the local portion, and the option to reach for the cloud only when you say so.
Comments