Perplexity Portable Computer gives teams a way to run powerful AI agents entirely on their own hardware, eliminating per‑token fees for any work that stays local. The system bundles a local model, inference engine, agent harness, tool sandbox and app connectors into one package that runs on an NVIDIA DGX Spark or any workstation with an RTX GPU offering at least 24 GB of VRAM. Every task begins on the device, so data never leaves the premises unless a user explicitly approves a step that needs live web access or frontier reasoning. Before any cloud call, the orchestrator shows exactly what would be sent, runs a PII classifier, and waits for a clear allow or deny decision. Approved steps go to one of 15+ cloud models, which return only text guidance and never see local files, tools or conversation history.
This design solves the main pain points for organizations that must keep sensitive information under strict control: finance teams reviewing fee disclosures, legal departments conducting PII‑bound research, healthcare groups handling patient data, government and defense units subject to residency rules, and engineering firms protecting intellectual property. Typical workloads—such as scanning dozens of brokerage PDFs for avoidable fees, migrating large code repositories, summarizing internal document sets, or triaging pull requests and posting approved summaries to Slack—can be completed fully on‑premises at zero marginal cost. When a step truly requires up‑to‑date web knowledge or a larger reasoning model, the user can opt‑in with full visibility, keeping spending predictable and avoiding surprise token bills.
Hardware requirements are clear: a DGX Spark with the GB10 superchip, 128 GB unified memory and ≥1 TB storage, or a Linux box with DGX OS/Ubuntu and an RTX GPU of 24 GB+ VRAM. Windows support arrives in September; macOS is not planned. The local orchestrator uses 3‑bit quantized Qwen 3.8 27B (≈17.4 GB download, 32 GB RAM) or the upcoming Nemotron 3.5 Lightning (4‑bit, ≈19 GB, 36 GB RAM). By keeping the core agent stack on‑device and gating any cloud use, Portable Computer delivers strong performance—85.4% on Perplexity’s internal knowledge work benchmark with its PPLX 27B model—while letting teams escalate only when they choose, and only after seeing exactly what will leave their machine.
#AI #Product #Enterprise #DataPrivacy #LocalAI #NVIDIA