Perplexity launched Portable Computer on August 25, a local AI agent product running entirely on NVIDIA DGX Spark hardware with no cloud dependency.
The full pipeline, including orchestrator LLM, subagent LLM, and agent harness, runs on-device using a 27B parameter model, meaning users pay no per-token costs. Perplexity said its harness scores 82.6% on real knowledge work benchmarks, outperforming open-source harnesses Pi and Hermes, and that a post-trained PPLX 27B model reaches 85.4% on the same benchmarks.