Perplexity launches local agent for DGX Spark workstation

Perplexity has released Portable Computer, a local agent application built specifically for Nvidia's DGX Spark workstation. The app runs a powerful personal agent on local hardware without sending data to the cloud or consuming remote model token quotas. The move extends the push toward agentic capabilities at the edge — in this case, a desktop box delivering 1,000 TOPS of AI compute.

Portable Computer connects directly to Google Drive, Gmail, Slack and GitHub, letting users switch with a click between a cloud model for complex tasks and a local model for routine work. When everything runs locally, Perplexity provides a one-click inference installation and an agent experience tuned to Nvidia hardware. The chosen local model is Qwen 3.8 27B in a version that has undergone dedicated post-training; a tuned variant of Nemotron 3.5 Lightning for especially fast responses is expected later. Windows and DGX Station support are in development.

In a parallel announcement, Nvidia said its RTX cards — including the RTX PRO series and the GeForce RTX 5090 — support Qwen3.8-27B from day one. The 27-billion-parameter open model was developed as a local companion to Qwen3.8-Max. It is optimized for coding with local files, tools and full project context, and reaches 131 tokens per second on a single RTX 5090 thanks to multi-token prediction (MTP). That performance enables extended development loops without moving sensitive code off the machine.

The model is already supported in llama.cpp, Ollama, Unsloth and LM Studio Bionic, so developers can start working with it immediately in familiar tools. Support also extends to DGX Spark and DGX Station, and Nvidia promises NVFP4 optimizations for all its local AI devices soon. The combination of RTX performance, the CUDA stack and open-source tooling provides a broad foundation for rapid adoption of new models.

GPUs remain. The Windows version of Portable Computer is not yet available, and DGX Station support is still on the way. Perplexity has not published comparative benchmarks against equivalent cloud models, and the tuned Nemotron 3.5 Lightning version has not been released. For now the launch targets early adopters willing to work on Linux and experiment with a full local agent, not a finished product for every user.