Perplexity introduces Portable Computer, separating AI reasoning from deterministic execution

Portable Computer Release

Perplexity has launched Portable Computer, a local-first version of its Computer agent. This new offering runs on an Nvidia DGX Spark workstation, with the base unit starting at $4,700. The architecture of Portable Computer introduces a significant shift in how AI agents are built for reliability.

Architectural Separation

Instead of relying on larger orchestrator, planner, or critic models for reliability, Perplexity's Portable Computer separates probabilistic reasoning from deterministic software. The local AI model proposes the next action, while deterministic code, referred to as the 'orchestrator' or runtime controller, assembles context, enforces policy, and executes approved tool calls within an OS-level sandbox. This design ensures that critical decisions are handled by inspectable software rather than solely by probabilistic AI.

Performance Improvements

Perplexity conducted tests using the Qwen3.8-27B model on the same DGX Spark across three different agent stacks. On its internal Local Knowledge Work Bench, a set of 53 tasks, Computer scored 82.6%, outperforming Pi (77.6%) and Hermes (74%). The performance gap widened on ParseBench-100, a subset covering charts, layouts, tables, and formatting, where Computer achieved 65.1% compared to Hermes's 34.6% and Pi's 13.9%. These results indicate that the system's architecture, rather than just the base model weights, significantly contributes to performance.

Implications for Enterprises

This approach, where a deterministic harness manages the control flow and policy enforcement, allows for more reliable and predictable AI agent behavior. Nate Kupp, Perplexity’s vice president of Computer Enterprise and Infrastructure, noted that the engineering effort was largely focused on developing this harness. This separation of concerns could be beneficial for enterprises seeking greater control and transparency over their AI agent deployments.