Perplexity launches Portable Computer for Windows RTX PCs

• Portable Computer is now in Perplexity’s Windows app on eligible NVIDIA RTX systems, with a 24GB-VRAM floor for on-device inference, according to Perplexity’s announcement.
• The local runtime keeps the model, agent harness, orchestrator, scheduler, files, and connected-tool activity on the PC, while Perplexity says a task can use cloud models when needed.
• Local MCP servers and recurring scheduled tasks arrive together, as Perplexity’s launch thread describes.
• Pro and Max plans have access under Perplexity’s Windows release post; testingcatalog’s reply says local tasks do not consume credits, while cloud-model calls still do.

Perplexity says a local task can pull in Search results or one of 15-plus frontier models, but it asks permission before transmitting work off-device. The model picker documented by testingcatalog’s inventory lists PPLX 27B and marks Nemotron 3.5 Lightning as coming soon.

Perplexity first put Portable Computer on NVIDIA DGX Spark in its August introduction. The Windows release moves the model, agent harness, orchestrator, and scheduler onto a compatible PC, according to the company’s Windows launch post.

The company says files, queries, and agent activity stay on the device during local runs.

The scheduler can run recurring analyses or file-processing jobs on the machine. Perplexity’s example is a daily freight-invoice reconciliation against rate sheets in a permitted folder, with an exception queue for duplicates and rate deviations.

Local MCP servers extend that workflow into desktop applications. Perplexity says the local model can use an application’s tools and data while keeping the model, source files, and interactions on-device; its allowed folders can hold documents, spreadsheets, presentations, PDFs, code, and images.

The same announcement shows a task being scheduled from the Windows interface with a local agent in Perplexity’s launch thread.

Portable can add cited web results through Perplexity Search or use more than 15 frontier models for advanced reasoning, the Windows release post says. It requests permission before moving work from the device to the cloud.

The Windows build supports NVIDIA GeForce RTX and RTX PRO GPUs only when they provide at least 24GB of VRAM. Perplexity says users choose a local model from the dropdown and download it in one click through the existing Windows app.

Testingcatalog’s reported model inventory names:

• PPLX 27B, described as Perplexity’s post-trained Qwen 3.8 27B.
• No stock Qwen 3.8 27B option on Windows.
• Nemotron 3.5 Lightning, described as roughly 30B MoE parameters with roughly 3B active, marked coming soon.