Perplexity's Portable Computer: A Fully Local AI Agent on NVIDIA Hardware

On August 25, 2026, Perplexity launched Portable Computer: a version of its autonomous agent platform that runs entirely on local hardware — no cloud dependency for the agent loop itself. It ships first on NVIDIA's DGX Spark desktop AI computer, and it represents the most aggressive attempt yet to move serious agent workloads off the cloud and onto your own machine.

The timing is not accidental. Agent products have a cost problem — long sessions burn tokens continuously — and a privacy problem — your confidential files have to leave your machine. Portable Computer attacks both at once: on-device work costs nothing per-token, and private data stays local.

What actually runs locally

This is the part most coverage glosses over. Portable Computer isn't a thin client with a local model bolted on. The entire agent stack runs on the device:

The models on offer are Qwen 3.8 27B and PPLX 27B — Perplexity's own post-trained variant of the same base model, tuned to complete as much of each task locally as possible. NVIDIA's Nemotron 3.5 Lightning is coming soon to the model picker, and bring-your-own-model configurations are supported.

If you've been watching the open-weights space, that model choice is the interesting detail: a post-trained Qwen 3.8 27B is now considered strong enough to power a serious always-on agent. For how to run that same model family on your own hardware, see our guide to running Qwen 3.8 27B locally.

Local-first, not local-only: the escalation model

Portable Computer's core design pattern is escalation. Every task starts on the device. When a step genuinely needs something the local model can't provide — live web access, or frontier-grade reasoning — the orchestrator pauses and asks permission:

1. The harness selects only the relevant context for that step
2. A PII classifier scrubs it
3. The user sees exactly what would be shared before anything leaves the machine
4. The approved step routes to one of 15+ cloud frontier models
5. The result comes back as guidance — the cloud model never gains direct access to your local files, tools, or the rest of the conversation

This is a clean answer to the classic privacy objection: the confidential details of a term sheet stay on the device, while the request for current market comps goes to the cloud.

Zero token economics

The unit economics are the real story for anyone building agent products. Work handled locally carries no per-credit charge. On Perplexity's own 53-task Local Knowledge Work Bench, PPLX 27B on a DGX Spark scored 85.4%. On hybrid benchmarks, fully-local runs scored below top cloud models — but closed most of the gap once cloud escalation was enabled, at a fraction of the per-task cost.

For the broader framework of how agent costs scale with tasks, rounds, and model choice, see our AI agent cost calculator and cost control guide.

The hardware requirement is the catch, of course. The launch configuration is the DGX Spark (GB10 superchip, 20-core Arm CPU, 128GB unified memory), plus Linux machines with NVIDIA RTX GPUs and 24GB+ VRAM. Notably absent from the roadmap: Apple Silicon and macOS. Local-first agents are, for now, an NVIDIA-hardware story — which explains why NVIDIA is co-marketing it so hard.

Security posture

Portable Computer inherits the sandboxing model from the cloud version of Perplexity Computer: code and tool execution run in OS-enforced sandboxes that restrict processes, file paths, and network access. The design principle worth noting: if the sandbox is unavailable, tool execution is disabled rather than allowed to run unchecked. That's the correct failure mode, and it's the opposite of what happened in the OpenAI sandbox-escape incident that dominated agent-security headlines this month — see how AI agents escape sandboxes.

App connectors (Google Drive, Gmail, Slack, GitHub) work through the local orchestrator: the agent can triage overnight GitHub issues against bug reports in Gmail, then post a Slack summary — all initiated on device.

Why this launch matters more than one product

Three trends are converging, and Portable Computer sits at the intersection:

1. Open-weight models got good enough. A 27B post-trained model now clears the bar for daily knowledge work. That's a threshold event — local agents stopped being a hobbyist curiosity this year.
2. Agent costs forced the issue. When agents run for hours, cloud token costs become the dominant expense. Local execution with selective escalation is the obvious architecture. NVIDIA's own research showed simple linear math can make model handoffs 25x faster in long agent sessions — the efficiency pressure is coming from every direction.
3. The hardware vendors need it. NVIDIA has been positioning DGX-class desktop machines as personal AI infrastructure. An always-on personal agent that costs nothing to run is the killer app for that hardware — especially now that NVIDIA is buying Hugging Face to own the open-source side of the ecosystem too.

Expect every serious agent vendor to ship a local-first mode within a year. The orchestrator/planner/router pattern here — with explicit escalation and data-minimization before anything leaves the device — is the template they'll copy.

Bottom line

Portable Computer is the first mainstream product to treat local execution as the default and the cloud as an accelerator, rather than the reverse. If you already own an RTX-class GPU, it's a working preview of how most agent workloads will run by 2027: private, cheap, and always on.

Related: Run Qwen 3.8 27B locally · AI agent cost control · AI agent security · NVIDIA buys Hugging Face · AI Agent Guide

Next Steps

See how the underlying local model family performs, and how to budget agent costs when you do escalate to cloud models.

how do AI agents work — return to the complete AI agent architecture guide.