Perplexity Launches Hybrid AI Computing for Cloud Intelligence and On-Device Privacy

Perplexity has launched hybrid computing for its AI agent platform. The new system enables a single AI agent to divide work between a frontier model operating in the cloud and a smaller open-weight model running locally on an Apple Silicon Mac. Sensitive information can remain on the user’s device instead of being transmitted to the cloud.

According to Perplexity, this is the first time an AI agent can begin a task in the cloud and dynamically transfer sensitive portions of that same task to a local model without restarting the workflow or losing context. The feature is available starting today through the Perplexity Computer platform.

Perplexity Pro and Max subscribers using Apple Silicon Macs with macOS 15 or later are eligible. Enterprise customers can also access the feature by opting in. The company says hybrid AI computing is designed to combine the reasoning capabilities of cloud-based frontier models with the privacy and security advantages of local execution.

“Hybrid is very attractive because often the most important thing to do when it comes to sensitive work is getting it right, so accuracy is very important,” Jon Staff, who leads Perplexity’s macOS and iOS engineering teams, said at a press conference attended by VentureBeat. “By combining the two, you can get maximum intelligence from the frontier model, but also get the security and privacy that comes locally.”

How Perplexity’s on-device privacy gate protects sensitive data

Perplexity’s hybrid architecture functions as an intelligent dispatcher. The cloud-based frontier model analyzes a user’s request, divides it into subtasks, and determines where each part should run. Web research, long-term planning, and complex reasoning can take place in the cloud, while private files, local data, and on-device actions are delegated to subagents running directly on the Mac.

The system’s central security feature is what Perplexity calls a privacy gate. This company-trained classifier operates on the user’s device and scans information for personally identifiable data, including names, addresses, account numbers, and other confidential details, before content is sent to the cloud.

If the privacy gate identifies potentially sensitive information, the user can decide whether to keep that portion of the task on the device or share it with the cloud model. This gives users greater control over how their data is processed.

“What we wanted to do is make sure that everything shared to the cloud orchestrator was secure,” a Perplexity representative said. “We built and trained our own PII classifier that integrates directly into Mac apps.”

The handoff between cloud and local models is managed automatically. Cloud orchestration breaks a prompt into separate tasks and routes each one to the appropriate subagent. When a task is assigned to a local model, it runs entirely on the Mac, and the related tokens are not sent to Perplexity’s cloud infrastructure.

Local processing may also reduce AI operating costs for organizations that track usage through cloud credits. Locally generated tokens do not consume those credits. Users remain responsible for the electricity and hardware costs, while credits are used for cloud-based orchestration and delegation.

How hybrid AI computing works for lawyers, investors, and business owners

Perplexity demonstrated its hybrid AI system with professional workflows involving confidential information that many organizations would avoid sending to a cloud-only agent.

In the first example, the cloud-based agent researched publicly available case law and submitted only anonymized legal questions. At the same time, a lawyer used a local subagent to update draft briefs against a privileged case file stored on a Mac. Perplexity said the privileged information remained on the device throughout the process.

The second demonstration focused on private equity. A group of agents reworked a financial model using confidential management projections, compared the transaction with publicly available market data, and produced a new version of an investment committee document. The process ran in the background for approximately 40 minutes without human intervention. According to Perplexity, manually combining local spreadsheets with cloud-based research would have taken several hours.

The third demonstration highlighted continuity across devices. A business owner began marketing research on an iPhone while traveling in an Uber. With permission, Perplexity Computer connected to the studio’s Macs and launched a local subagent to analyze customer interviews and revenue data. The system then combined those private findings with cloud-based research into competitors and publicly listed prices.

“It doesn’t matter how far away she is from the computer,” the Perplexity representative said.

The company argues that workflows of this type are difficult to support in either a completely local or completely cloud-based environment. Fully local systems may lack advanced reasoning capabilities, while cloud-only agents can create unacceptable privacy and compliance risks.

“Tasks like this are not possible in a fully local or fully cloud environment,” the representative said. “We need local security and privacy, but we also need frontier intelligence.”

Why Perplexity’s use of local Qwen models is drawing attention

At launch, Perplexity users can select from three local models, including Google Gemma E4B, Alibaba’s Qwen3.6 35B-A3B, and a post-trained version of the Qwen 3.6 35B model. Perplexity recommends these models for local execution on compatible Apple Silicon hardware.

The inclusion of a model developed in China raises questions for enterprise and government customers that have strict requirements around software supply chains and data sovereignty. Perplexity says those concerns are reduced because the models run locally and do not transmit user data to an overseas cloud provider.

“The great thing about these models is that they’re indiscriminate weights. We can evaluate them ourselves,” the representative said. “If that model is running locally on your computer, the data doesn’t leave the computer itself. We don’t actually send those tokens to a cloud provider that’s hosted in another country.”

Perplexity also says its models are hosted in the United States when cloud processing is required. In addition, the local agents operate inside the macOS sandbox framework, commonly associated with Seatbelt, which limits what an agent can access or change on the computer.

If a local agent attempts an action outside its permitted capabilities, the system can stop the action and request user approval. Perplexity does not currently offer unrestricted “YOLO mode” execution, although the company suggested that broader controls could eventually become available for certain users.

Enterprise administrators can establish organization-wide confidentiality policies and maintain audit records showing what information leaves each device. These controls are intended for industries such as law, finance, healthcare, and other sectors with strict compliance obligations.

Questions remain about how consumer usage data may be handled for product improvement and model training. Perplexity pointed to its incognito mode and existing opt-out settings, while enterprise agreements can include zero-data-retention provisions. The company said it does not use its systems for post-training applications globally and indicated that it would provide additional details about nonenterprise accounts.

Hybrid AI could address enterprise privacy and data governance challenges

Perplexity’s announcement arrives as businesses continue to struggle with privacy, security, and data governance in generative AI deployments. Much of the most valuable corporate work involves confidential information that organizations are unwilling to place on third-party servers.

The NIST Generative AI Risk Profile identifies data privacy and information leakage as key risks associated with generative AI. McKinsey research on the state of AI has also found that data governance remains a major obstacle as companies move from experimentation to wider business adoption.

Gartner’s hybrid computing trend similarly reflects growing interest in architectures that distribute workloads across local devices, private infrastructure, and public cloud platforms.

Perplexity is positioning its hybrid AI platform as an alternative to choosing between cloud intelligence and local privacy. Its orchestration layer acts as a real-time mediator, deciding which parts of a task should be processed by a frontier cloud model and which should remain on the user’s hardware.

The strategy also supports Perplexity’s broader position as an application layer that works with multiple AI models. As local models improve and Apple Silicon hardware becomes more capable, the company can potentially improve the overall experience without relying on a single model provider.

“Every time one of these things improves, so does Perplexity,” the representative said. “That’s the really cool nature of us being in this application layer and coordinating all the different pieces together.”

Perplexity’s rapid growth from a $520 million startup to a major AI platform

Hybrid computing expands Perplexity’s broader effort to make AI agents an interface for completing tasks rather than simply answering questions in a chat window.

The company introduced Comet AI Browser in July 2025 as an early attempt to turn agents into a primary interface for computing. The browser initially launched with a maximum subscriber price of $200 per month. Perplexity Computer followed in March 2026, along with desktop applications for Mac and Windows.

More recently, the company introduced a local-first version of Computer and support for NVIDIA DGX Spark hardware. That approach begins processing on the user’s device and escalates to a cloud model only with permission. The latest release reverses that workflow by starting in the cloud and delegating sensitive tasks down to local models.

Perplexity’s business trajectory has been similarly ambitious. The company was valued at $520 million in January 2024 and reportedly reached a $20 billion valuation by September 2025.

Along the way, Perplexity made a reported $34.5 billion bid for Google’s Chrome browser during the debate over Google’s antitrust remedies. Bloomberg also reported that Apple executives had held internal discussions about potentially acquiring the AI startup.

However, the company faces significant challenges. Reuters reported that a Reddit data-scraping lawsuit against Perplexity survived a motion to dismiss. The company is also facing broader copyright and data-access disputes across the AI industry.

Hardware requirements could limit adoption of Perplexity’s local models. The company recommends at least 32GB of unified memory for stronger local performance, while smaller configurations may deliver significantly lower performance. Support for Windows and Linux is expected to arrive later, meaning the initial release is focused heavily on Apple Silicon Macs.

The privacy gate itself also presents a potential weakness. Because it is a machine-learning classifier, it may fail to identify certain sensitive information. A false negative could result in confidential data being sent to the cloud. Perplexity’s approach is to give users visibility into what the gate detects and provide businesses with device-level audit logs.

Ultimately, Perplexity is asking users and organizations to trust an AI system to determine which data it can process and where that processing should occur. Its central argument is that enterprise AI privacy may not require completely separate local and cloud systems. Instead, the company believes smarter privacy controls and hybrid AI orchestration can deliver both advanced reasoning and stronger data protection.