Perplexity Runs Its Computer Agent on Local Hardware
Portable Computer keeps agent planning, file analysis and tool execution on a DGX Spark, escalating to cloud models only with permission.
An agent stack contained on the desk
Perplexity has launched Portable Computer, a local-first version of its Computer agent designed initially for Nvidia's DGX Spark. Unlike a desktop interface that merely forwards requests to a cloud model, the product runs its orchestrator, planner, tool router, scheduler, durable task queue and local search index on the user's machine.
Users can choose Qwen 3.8 27B or PPLX 27B, Perplexity's task-specific post-training of that model. Nemotron 3.5 Lightning is due to join the model picker later. The agent can read and search local documents and code, create files, execute tools and continue long-running jobs without charging credits for locally completed steps. Voice transcription also stays on the device through Nvidia's Nemotron speech-recognition model.
Portable Computer can connect to Gmail, Google Drive, Slack and GitHub. When a workflow needs current web information, a connected application or stronger reasoning, it may request permission to consult Perplexity search or more than 15 cloud models. Perplexity says transfers from the device require user approval and that code execution occurs inside isolated sandboxes with restricted file and application access.
Privacy has a hardware threshold
The initial release is available to Perplexity Pro and Max subscribers using Linux on DGX Spark, whose Grace Blackwell GB10 platform includes 128 GB of unified memory. Windows and RTX PC support are promised but are not yet available. The requirement substantially narrows the launch audience and means the product has not yet established that this architecture works conveniently on ordinary workplace computers.
Still, Portable Computer is a consequential test of hybrid agents. It divides a task at execution time rather than forcing customers to choose permanently between private but weaker local software and capable but remote cloud services. If its permission boundaries are dependable, organizations could keep confidential source code, contracts and internal correspondence on premises while purchasing frontier reasoning only for selected steps. Its benchmark claims and security controls now need independent testing, particularly around what context an escalation request exposes.