Perplexity introduces portable computer for local AI workflows on nvidia dgx spark

Perplexity has launched Portable Computer, a locally-first version of its agentic AI platform designed to run on NVIDIA DGX Spark systems. The system integrates the agent harness, orchestrator, planner, tool router, and post-trained models directly on the device, eliminating the need for cloud-based processing for local tasks. This setup ensures that work handled by local models incurs no per-token cost, making it economically viable for large-scale operations on owned hardware. The system includes an OS-enforced sandbox that restricts process execution, filesystem access, and network activity, ensuring security and isolation. If the sandbox is unavailable, tool execution is disabled entirely rather than being compromised. Integration with services like Gmail, Outlook, Slack, and GitHub is facilitated through the local orchestrator. Portable Computer supports both Qwen 3.8 27B and PPLX 27B models, with the latter being specifically tuned for Perplexity’s harness. It also allows users to bring their own models and inference servers. The system is optimized to work around small-model context limitations by maintaining compact system prompts, loading specialized skills on demand, and using compact CLI tools instead of full MCP definitions. Performance benchmarks on Perplexity’s 53-task Local Knowledge Work Bench showed that Computer running Qwen 3.8 27B achieved 82.6%, outperforming the open-source Pi harness at 77.6% and Hermes at 74.0%. PPLX 27B further improved results to 85.4%. On BrowseComp, Computer scored 66.7% compared to 50.2% (Pi) and 43.9% (Hermes), using 51% less wall time and 70% fewer tokens. On ParseBench-100, it reached 65.1% against 34.6% and 13.9%. The hybrid approach, combining local execution with cloud escalation, demonstrated effectiveness. On Terminal Bench 2.1, a fully local run scored 59.6% at near-zero cost, while escalation to cloud models increased performance to 73.0% at approximately $0.415 per rollout. This compares to 82.4% at about $0.65 for Claude Opus 5 alone. Escalation narrows the gap with frontier models but does not fully close it. DGX Spark installation requires a GB10 superchip, 128 GB of memory, and at least 1 TB of storage. Qwen 3.8 27B is available at 3-bit quantization, requiring 32 GB RAM, while Nemotron 3.5 Lightning is 4-bit, needing 36 GB. Other systems must use DGX OS or Ubuntu on ARM or x64 with an RTX GPU with 24 GB or more VRAM. Installation is done via an apt repository. Availability is initially limited to Linux for Pro, Max, Enterprise Pro, and Enterprise Max subscribers, with Windows support coming in September and macOS not on the roadmap. Only one DGX Spark is supported at launch, with clustering planned for the future. The zero per-token cost for local steps is a key factor in enabling large-scale migrations and long verification loops on owned hardware.