Perplexity Brings Local AI Agents to Windows, But GPU Requirements Remain Steep

Perplexity has extended its Portable Computer offering to Windows machines equipped with NVIDIA RTX and RTX PRO graphics cards, delivering a complete local AI agent environment that combines model runtime, orchestration, security and hardware integration. The challenge of deploying an LLM locally differs significantly from deploying an autonomous agent, and Perplexity's approach bundles these components together rather than leaving them separate. Access requires an NVIDIA GPU with a minimum of 24GB of VRAM.

The Windows release marks Perplexity's third platform launch in under three weeks. Portable Computer debuted on Linux and NVIDIA DGX Spark on August 25, followed by a hybrid compute option for Apple silicon a week later, which distributes processing between local and cloud models on Mac systems. Bringing the product to Windows demanded more than a straightforward port; Perplexity adapted the model runtime, orchestration, security mechanisms and hardware integration layers for each platform while maintaining consistent user experience across all three.

Orchestration beyond the model

Portable Computer packages orchestration, tool calling and sandboxing alongside the model itself. On Windows, the application supports PPLX 27B—Perplexity's post-trained model—and Qwen 3.8 27B, both tuned for RTX GPUs, along with an integrated browser, tool calling capabilities, and Perplexity's proprietary SPACE sandbox.

This positioning differs from tools like LM Studio or Ollama, which prioritize simplicity in running models locally but do not extend to granting models autonomy over complex, multistep tasks. DeepSeek's recent hiring wave of approximately 150 positions, predominantly centered on agent infrastructure rather than model development, underscores the engineering complexity separating a strong model from a functional agent. As one observer noted, "DeepSeek's recent hiring spree of roughly 150 new roles, nearly all of them focused on agent infrastructure rather than the model, hints at how much engineering sits between a capable model and a capable agent."

Connectors blur local boundaries

Perplexity includes connectors for Microsoft Outlook, OneDrive, and Word, as well as Google Drive, Gmail, Slack, and GitHub. This roster reveals what the term "local" encompasses in this context.

The agent reaches external services because it operates without air-gapping. Tasks completed locally can process files without transmitting documents to a cloud-based model. Once an agent gains access to both local files and remote APIs on the same machine, determining which resources the agent actually requires and where to locate them becomes more complex.

Hybrid cloud as fallback

Perplexity acknowledges that a 27-billion-parameter model operating on a desktop GPU cannot handle all scenarios, which explains the hybrid design. When the agent recognizes that a task exceeds the local model's reasoning capacity, it can route the request to Perplexity's cloud infrastructure.

NVIDIA states that the agent determines when cloud assistance would be beneficial and requests user permission before transferring any data outside the machine.

For enterprises managing sensitive or regulated information, this division carries significant implications. A local agent can process source code or financial records without uploading them to a hosted model for routine operations. The economics matter as well: tasks handled locally do not consume Perplexity Computer credits.

High VRAM floor limits reach

Portable Computer is accessible through Perplexity Pro ($20/month) and Max ($200/month) subscription tiers, available for both individual and enterprise customers, with NVIDIA DGX Station support arriving at a later date. The broader challenge involves transitioning local agents from experimental developer projects into production-ready enterprise solutions. By integrating this capability into Windows, Perplexity immediately positions itself in front of the user base scale necessary to achieve that transition.