Perplexity AI Inc. has launched Hybrid Compute, a new platform designed to dynamically route AI tasks between cloud-based servers and local on-device processing. The launch continues a rapid succession of agentic AI tool releases from the company. CEO Aravind Srinivas announced the software during the Computex conference in Taipei alongside Intel.

Srinivas described the system as acting like an "air-traffic controller" for AI tasks. It decides in real time which jobs can run locally on a personal computer and which require more powerful cloud servers. The system aims to manage the massive demand for artificial intelligence computing power by easing the strain on centralized servers.

This release follows the company's debut of Perplexity Computer in February, Personal Computer in April, and Portable Computer late last month. Hybrid Compute is pitched as a middle ground between the Personal Computer and Portable Computer platforms. The company also detailed the new platform's mechanics and target audience in an announcement post on X.

Automatic Privacy Routing

Currently available in the Perplexity app for Apple Silicon Macs for Pro or Max subscribers, the system automatically detects sensitive information in user files. The earlier Perplexity Computer suite, similar to Claude Cowork, featured AI agents that could autonomously complete tasks using the web, PC files, and local apps. Hybrid Compute builds on this by securely routing confidential data to local models while sending complex reasoning to the cloud.

According to Jon Staff, who oversees Perplexity's Mac products, the app uses a newly trained privacy classifier to automatically identify sensitive content. Users can review these file suggestions before deciding how to split the task, double-checking to ensure nothing important was missed. They can then choose which specific models will handle the work, or they can simply upload everything to the cloud.

Perplexity suggests the system is highly beneficial for professional confidentiality. For instance, a lawyer preparing a brief that compares an active case against existing case law could use the tool to keep their client's data strictly local. This hybrid approach also helps thrifty users reduce overall inference costs, as processing handled by local models does not incur token charges.

Supported AI Models and Interface

Locally, the platform currently supports Gemma 4 E4B alongside two flavors of Qwen's 35-billion-parameter 3.6 model. One of these Qwen variants is a custom post-trained version developed by Perplexity. The company plans to add more local models in the future.

For cloud options, users can take advantage of frontier models like Anthropic's Claude Opus 5 and GPT 5.6 Sol. Offloading work from these typically pricier frontier models directly contributes to user savings.

The Mac application interface features a real-time visualization displaying local CPU, GPU, and memory usage. A sidebar sits alongside this data, tracking how many tokens the task has consumed.

Installation of local models is handled entirely within the app, completely bypassing the need for users to open or use Mac terminal commands. Users can write follow-up instructions once an output is generated and can even queue up tasks using an iPhone.

This product launch comes at a time of significant growth for the AI search company. Perplexity recently saw its revenue hit $500 million, even as its headcount grew by just 34%.