Perplexity Portable Computer: An Agent That Stays on Your GPU
Perplexity's Portable Computer (August 25) runs its full agent stack — orchestrator, planner, tool router, search index — locally on an Nvidia GPU box, with zero per-token cost for local steps and a per-step approval gate before anything reaches the cloud. We walk through the hardware bar, installation, the two local models, and what the escalation gate actually sends.
There's a new kind of "local AI" in town, and it is not a chat app with a file picker. On August 25, 2026, Perplexity and Nvidia shipped Portable Computer — the cloud Computer platform's full runtime (orchestrator LLM, sub-agent LLM, agent harness, planner, tool router, scheduler, persistent task queue, even the search index) rebuilt to run entirely on your own hardware. If you have a Pro, Max or Enterprise subscription, it's included at no extra cost. Here's what running it actually takes.
The Hardware Bar (Read This First)
Two paths. Reference device: the Nvidia DGX Spark — GB10 Grace Blackwell superchip, 20-core Arm CPU, 128 GB unified memory, DGX OS, 1 TB free storage. It started at $3,999 and has drifted to about $4,699 amid memory shortages. Your-own-box path: any x86 machine running Ubuntu (ARM works too) with an RTX GPU carrying 24 GB+ of VRAM — think RTX 3090, 4090, 5090; even a used 3090 runs $1,500+. Launch is Linux-only (DGX OS or Ubuntu); a Windows version lands in September. macOS is not on the roadmap, and only a single DGX Spark is supported for now (clustering is roadmap, not shipped).
Setup and the Two Models
Installation is a standard apt repository add (or one-click from the Perplexity app). You pick one of two local models: Qwen 3.8 27B (3-bit quantization, a 17.4 GB download, 32 GB RAM required) or PPLX 27B, Perplexity's own post-trained variant tuned for its harness; Nvidia's Nemotron 3.5 Lightning (30B MoE) is listed as coming soon, and bring-your-own model is supported. Inference runs on vLLM, with local speech recognition via Nemotron 3.5 ASR. Because Qwen 3.8 27B advertises a 260K-token window but degrades past ~100K, the harness keeps the system prompt tiny, loads skills on demand, exposes connectors as compact CLI tools instead of full MCP definitions, and compacts stale context mid-run.
The Escalation Gate: What Actually Leaves the Machine
Default posture: nothing leaves. Code and tool calls run inside an OS-enforced sandbox (processes, filesystem paths, network access), and if the sandbox is unavailable, tools are disabled rather than silently downgraded. Gmail, Outlook, Slack and GitHub connectors route through the local orchestrator. When a step genuinely needs the live web or frontier reasoning, the orchestrator stops and asks: it selects the relevant context, runs a PII classifier over it, and shows you exactly what would leave — then, only after approval, sends that single step to one of 15+ cloud models (GPT-5.5, Claude Fable 5 and friends). The remote adviser returns text guidance only; it never touches your files, tools or conversation.
Does It Work? The Numbers (Self-Reported)
On Perplexity's own 53-task Local Knowledge Work Bench (which it plans to open-source), Qwen 3.8 27B scored 82.6% versus 77.6% for the open Pi harness and 74.0% for Hermes on the same model; PPLX 27B reached 85.4%. On Terminal-Bench 2.1, a fully local run scored 59.6% at effectively zero marginal cost; adding adviser escalation lifted it to 73.0% at roughly $0.415 per rollout, against 82.4% at ~$0.65 for Claude Opus 5 alone. Escalation narrows the gap to frontier models without closing it. Caveat: every number above is Perplexity's own — independent reviews are still thin on the ground, and a box that runs 24/7 does draw real electricity.
Verdict: if you already own a 24 GB+ RTX card and a Perplexity subscription, this is the most complete local-first agent system you can run today — a full agent stack, not just a model runtime like Ollama. For everyone else, the machine is the price of entry, and the honest pitch is "your data never leaves without you watching it leave."