Perplexity has added a Windows client for its Portable Computer agent, allowing the model to run entirely on a local Nvidia GPU that meets a 24 GB VRAM minimum. This shift moves inference, tool‑calling, and sandboxed browsing from the cloud to the desktop, which directly impacts runtime configuration, cost accounting, and data‑handling policies for AI, platform, and security engineers.
Runtime and Hardware Requirements
The Windows build supports two post‑trained models – PPLX 27B and Qwen 3.8 27B – both tuned for RTX‑class GPUs. The agent bundles a built‑in browser, tool‑calling framework, and Perplexity’s proprietary SPACE sandbox. Because the runtime is bound to a single GPU, the only supported hardware is an Nvidia GPU with at least 24 GB of VRAM. Engineers must verify that target machines meet this threshold before deployment, and plan for the associated cost of high‑end GPUs or DGX‑type stations for enterprise scaling.
Hybrid Cloud Fallback and Data Handling
Perplexity acknowledges that a 27‑billion‑parameter model on a desktop cannot cover every workload. When the local agent detects a task that exceeds its capacity, it can hand off processing to Perplexity’s cloud models. The handoff is user‑prompted, and data is only transmitted after explicit permission. This design gives organizations a way to keep sensitive files – such as source code or financial records – on‑premises for routine work while still accessing larger models when needed, reducing cloud credit consumption for local‑friendly tasks.
Integration, Orchestration, and Connector Landscape
Beyond the core model, the agent ships connectors for Microsoft Outlook, OneDrive, Word, Google Drive, Gmail, Slack, and GitHub. These integrations mean the agent can act on both local files and remote APIs from the same machine, blurring the line between “local” and “cloud”. Orchestration on Windows required adapting the model runtime, security sandbox, and hardware integration layers to keep the user experience consistent with the earlier Linux and Apple‑silicon releases.
Operational and Security Considerations
- Resource Planning: The 24 GB VRAM floor limits deployment to high‑end workstations or specialized hardware, influencing capacity planning and budgeting.
- Cost Model: Local execution avoids Perplexity Computer credits, but the hardware investment and Perplexity Pro ($20 / month) or Max ($200 / month) subscriptions add recurring costs.
- Data Residency: Sensitive data can stay on the device for tasks handled locally, but any escalation to the cloud requires user consent, which must be incorporated into UI/UX and policy enforcement.
- Security Surface: The agent’s ability to reach external services via connectors expands the attack surface; security teams should assess connector permissions and sandbox isolation provided by SPACE.
- Version Management: Keeping the Windows client in sync with Linux and Apple releases may require coordinated updates to runtime libraries and model binaries.
Related CloudNinjas coverage: AI engineering.
What This Means For Practitioners
Engineers should evaluate whether their workloads justify the GPU investment and whether the hybrid fallback aligns with data‑privacy requirements. Validate connector configurations and sandbox behavior before granting the agent access to production resources. Monitor credit usage to understand the cost trade‑off between local execution and cloud escalation. Finally, incorporate the 24 GB VRAM prerequisite into hardware procurement standards to avoid deployment surprises.
