AI
Perplexity brings local AI agent to Windows—but demands 24GB GPU memory
Portable Computer, an agent that runs entirely on your GPU, is now available on Windows PCs with compatible Nvidia GPUs, though high memory requirements may limit adoption.
The New Stack
Perplexity has launched Portable Computer for Windows, allowing its AI agent to operate locally on compatible Nvidia GeForce RTX and RTX PRO GPUs. The catch is substantial: you'll need at least 24GB of VRAM to run it.
The Windows rollout represents the third major platform for Portable Computer in less than three weeks, following releases on Linux and Nvidia DGX Spark in late August, plus hybrid compute capabilities for Apple silicon. However, bringing the technology to Windows required significant adaptation beyond a simple port—Perplexity reworked the model runtime, orchestration, security layer, and hardware integration to maintain a consistent user experience across platforms.
What Makes It More Than Just a Local Model#
Unlike lighter-weight tools like LM Studio or Ollama that simply make model inference accessible, Portable Computer functions as a complete agentic system. It includes PPLX 27B and Qwen 3.8 27B models (both tuned for RTX GPUs), an integrated browser, tool calling capabilities, and Perplexity's proprietary SPACE sandbox environment.
The distinction matters. Running a model locally is straightforward; orchestrating that model to autonomously manage multistep workflows is fundamentally different engineering work. DeepSeek's recent recruitment of roughly 150 roles—nearly all focused on agent infrastructure rather than model development—underscores how much invisible complexity separates a capable model from a capable agent.
Hybrid Cloud as a Safety Net#
Perplexity acknowledges that a 27-billion-parameter model on a desktop GPU cannot handle every scenario. The system uses a hybrid architecture where the agent can escalate complex reasoning tasks to Perplexity's cloud models when needed. According to Nvidia, the agent determines when cloud assistance would help and requests user permission before transmitting any data off the machine—a feature that appeals to organizations handling sensitive or regulated information. Tasks completed locally also avoid burning Perplexity Computer credits.
Connectors and the Blurred "Local" Boundary#
Portable Computer ships with connectors for Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub. This integrated access suggests that "running locally" doesn't mean air-gapped isolation. The agent can process files on your machine without uploading them to cloud inference services while simultaneously reaching external APIs and remote resources. As access to both local and remote systems converges on a single machine, determining which resources an agent should actually use becomes increasingly complex.

