Perplexity brings its AI agent to your desktop—for a steep price

Perplexity and Nvidia have launched Portable Computer, a locally-run version of Perplexity's agentic AI assistant that brings sophisticated task-handling capabilities to desktop machines. The move reflects growing demand for on-device AI, though the hardware requirements remain a significant barrier to adoption.

Hardware hurdles and affordability#

Running Portable Computer demands serious equipment. Users can either purchase Nvidia's DGX Spark desktop—priced at $4,800—or equip an existing Ubuntu machine (ARM or x64) with an Nvidia RTX GPU boasting at least 24GB of memory. Even second-hand RTX 3090 cards, which meet the minimum specs, currently sell for over $1,500. These costs are unlikely to drop soon, as the ongoing global memory shortage continues to tighten supply.

The engineering team had to rebuild the system from the ground up to make the cloud-based Computer agent work on constrained local hardware. According to Nate Kupp, VP of Computer Enterprise and Infrastructure, the company "revisited almost everything throughout the stack." While many of Computer's underlying capabilities remained intact, the orchestration framework and model configuration required substantial redesign to function efficiently with fewer parameters than Perplexity's larger cloud models.

The local orchestration architecture#

The heart of Portable Computer is its deterministic orchestrator—not an AI model itself, but executable code that manages the agent's decision loop. When the local model (currently based on Qwen 3.8-27B or Perplexity's own fine-tuned variant) proposes an action, the orchestrator assembles context, applies security policies, and executes approved tool calls within an OS-level sandbox.

This sandbox proves essential. It restricts process spawning, filesystem access, and network communications. If the sandbox becomes unavailable, the harness shuts down rather than risk running tools unsandboxed. The design also adapts to practical limits: while the Qwen model supports a 260,000-token context window, performance degrades beyond 100,000 tokens, so Portable Computer keeps the core prompt lean and loads additional capabilities on demand.

Rather than expose full Model Context Protocol definitions directly to the model, frequently-used connectors are now command-line tools, reducing the overhead the model must manage.

Performance and benchmarks#

Testing reveals the engineering investment paid off. On Perplexity's internal 53-task Local Knowledge Work Bench, Portable Computer scored 82.6%—ahead of competitors Pi (77.6%) and Hermes (74%). On ParseBench-100, a suite covering charts, layouts, tables, text, and formatting, it scored 65.1%, substantially outpacing Hermes (34.6%) and Pi (13.9%). The company plans to open-source these benchmarks.