Yesterday Perplexity announced Portable Computer on the NVIDIA DGX Spark:
Portable Computer is a fully local version of Perplexity Computer, where the entire runtime: orchestrator LLM, subagent LLM, agent harness all run on your local hardware. No cloud dependency.
The post gives the basic configuration: Qwen 3.8 27B or PPLX 27B (a post-trained version of that model), in 4-bit quantization, on a GB10 box with a 20-core Arm CPU and 128 GB of unified memory. The orchestrator, planner, tool router, scheduler, durable task queue, and local search index all live on the machine. Local work carries no per-credit charge. When a task needs the outside world — current information, browser use, frontier reasoning — the local orchestrator asks permission first, routes just that step to one of 15+ cloud models, and brings the answer back into the same run.
The boundary I care about
Downloadable weights are not the new part. Perplexity says the planner, tool router, scheduler, task queue, local search index, orchestrator, and subagents run on the DGX Spark. That moves the repeated coordination work off Perplexity's servers and removes the per-credit charge for work completed locally.
It does not mean every task or document stays on the device. Search, connected apps, and frontier models remain cloud services. When the local run needs one of them, Portable Computer asks for permission before escalating that step; information included in an approved call may then leave the machine. The permission prompt, rather than a promise of permanent isolation, is the useful boundary.
I also would not read local execution as a promise that the harness is user-modifiable. The announcement does not establish that. What it does establish is narrower: the agent loop can run locally, and the user approves cloud crossings.
This first release is a Linux workstation setup for the DGX Spark's 128 GB unified-memory class, not a laptop mode. Perplexity lists Windows as coming later. The test I would run is simple: give it a job over private local files, decline escalation, and see how far the 27B models get before the permission boundary becomes the bottleneck.