This collaboration marks a significant move to shift complex AI workloads—including data analysis and report generation—away from the cloud to local devices, ensuring that sensitive files and processing remain private.
The transition to local execution addresses the high costs and privacy concerns associated with "agentic" workloads, which involve AI performing multi-step tasks that consume vast amounts of data.
By running locally, the software incurs zero "token" fees—the standard usage costs typically paid to cloud AI providers.
Perplexity notes that while the system starts tasks on-device by default, it can escalate more difficult problems to powerful cloud-based models with the user's permission, offering a balance between local privacy and frontier-level performance.
To optimize performance on local hardware, Perplexity developed a specialized "agent harness"—the software framework that manages how the AI uses tools and processes information—specifically for smaller open-source models like Qwen 3.8 and its own PPLX 27B.
This approach reportedly uses significantly less memory and processing time than general-purpose frameworks.
While the initial release targets enterprise and power users on Linux, a Windows version is scheduled for release in September, further expanding the availability of local AI infrastructure.