Perplexity Portable Computer finally hit Windows for those of us with RTX GPUs

Local AI agents are finally moving past the "cool demo" phase and actually doing work on the machine. Perplexity just dropped Portable Computer for Windows, which is basically the local version of their agent that handles multi-step tasks without shipping every single byte of your data to a cloud server. The big draw here is that local work doesn't eat into your Perplexity Computer credits, and your sensitive files actually stay on your drive.

Does my hardware actually cut it?

If you're wondering why it's not launching, check your VRAM. This isn't for your old office laptop. You need an NVIDIA GeForce RTX or RTX PRO GPU with at least 24GB of VRAM to run this. It's already supported on Linux and NVIDIA DGX Spark systems, but the Windows app is the new addition.

How does it handle the "local vs cloud" split?

The app uses a local model—specifically Qwen 3.8 27B, which has been post-trained for the Perplexity Computer environment and optimized for RTX hardware. You don't have to spend three hours fighting with CUDA versions or configuring a software stack just to get a prompt to run.

It uses a proprietary SPACE sandbox and a built-in browser to execute tasks. If the local model hits a wall and needs more "brain power" for advanced reasoning, it'll actually ask for permission before it offloads the task to a cloud model.

What can this actually do besides summarize PDFs?

Since it has connectors for GitHub, Slack, Gmail, Google Drive, OneDrive, Word, and Outlook, it's actually useful for workflows that usually require a dozen open tabs:

• GitHub: It can scan open pull requests, sort them by status, and find outdated docs that need updating.
• Finance: You can feed it two years of 1099s and brokerage summaries to find tax drag and fees. The best part is the citations point to the exact page and file without the documents ever leaving the device.
• Ops: It can analyze funnel exports locally to find where users are dropping off and then dump those insights into Slack.

Other NVIDIA local AI updates to track

While I'm digging into the local agent stuff, there are a few other model updates happening in the NVIDIA ecosystem:

• GLM 5.3: Z.ai's 744B flagship model is tuned for long agent sessions on DGX Station and four-node DGX Spark clusters.
• GLM 5.3 Flash: Optimized for DGX Station and dual DGX Spark setups for lower-cost visual intelligence.
• DeepSeek-v4.1 Flash: This one is focused on reducing KV cache memory demands, which is a win for complex agent workloads.
• Qwen3.8-Flash-Next: A multimodal MoE model that's open-weight.
• Qwen4: There's an early preview that runs on a single DGX Spark using NVFP4.