Perplexity Ships Portable Computer on NVIDIA DGX Spark

Perplexity has released Portable Computer, a local-first version of its agentic computing platform designed to run its agent harness, orchestrator, planner, tool router, and post-trained models directly on NVIDIA DGX Spark hardware. This new offering packages the local model, inference engine, tool sandbox, and application connectors into a single system, allowing every task to commence on the user's device. Work handled by local models incurs no per-token charges, a significant cost-saving measure for extensive local processing. When a task requires access to the live web or advanced reasoning capabilities beyond the local model's capacity, the orchestrator pauses and requests explicit permission before transmitting the specific step to one of over 15 cloud-based models. The deployment of Portable Computer is contingent on specific hardware requirements, necessitating a GB10-class box or an RTX GPU with at least 24 GB of VRAM. This release represents shipping software rather than a preview binary. The target market for Portable Computer includes enterprises and mid-market teams that already possess NVIDIA workstations, as well as well-funded AI-native startups. It is not considered viable for general small and medium-sized businesses due to the substantial initial hardware investment. Industries that stand to benefit most include finance, legal, healthcare, government and defense, and intellectual property-heavy engineering sectors, particularly in environments where data residency or contractual confidentiality mandates prohibit cloud-based inference. Potential applications range from fee and disclosure reviews across large document sets, PII-bounded research, and repo-scale data migrations to batch summarization of local corpora and PR triage that integrates with platforms like Slack. Portable Computer is not merely a local chat application with a file selection feature. Instead, Perplexity integrates the local model, inference engine, agent harness, tool sandbox, and application connectors into a cohesive system, thereby eliminating the conventional complexities of setting up an inference server and manually configuring tools. Users can select between Qwen 3.8 27B and PPLX 27B, Perplexity’s proprietary post-trained variant optimized for its harness. Additionally, NVIDIA Nemotron 3.5 Lightning, an open 30B Mixture-of-Experts (MoE) model, is slated for future availability. The platform also supports bring-your-own-model and inference server configurations. All code and tool calls are executed within an operating system-enforced sandbox, which imposes strict limitations on processes, filesystem paths, and network access, ensuring a secure execution environment. If the sandbox environment is compromised or unavailable, tool execution is automatically disabled to maintain system integrity.