Perplexity Hybrid Compute Splits AI Tasks Between Cloud and Local Models to Protect Sensitive Data

Perplexity Introduces Hybrid Compute for Privacy-Focused AI Processing

Perplexity has launched Hybrid Compute, a feature that enables users to split tasks between cloud and local AI models to keep sensitive data secure on their devices1

1

2

2

. The new capability, available through Perplexity's Mac app, allows a single AI agent to dynamically divide work between cloud-based frontier models like Opus 5 or GPT-5.6 Sol and local LLMs running on Apple Silicon Macs4

4

. Jon Staff, who oversees Perplexity's Mac products, explained that the system automatically checks for sensitive content before uploading: "Any time you try to upload files or send information, we're going to automatically check for sensitive content and make sure that you want to share that data to the cloud"1

1

.

This marks the first time an AI agent can begin AI tasks in the cloud and dynamically hand off confidential portions to a model running on the user's own hardware without restarting the job or losing context4

4

. The feature becomes available today for enterprise customers, Pro subscribers, and Max subscribers on any Apple Silicon Macs running macOS 15 or later5

5

.

Privacy Classifier Gates Sensitive Data from Cloud Processing

At the core of Hybrid Compute sits a privacy classifier trained by Perplexity to automatically identify personally identifiable information (PII) including names, addresses, and account numbers before transmission to the cloud3

3

4

4

. The company has open-sourced this on-device PII classifier, which was trained in collaboration with Perplexity's Secure Intelligence Institute3

3

. When the Privacy Gate flags sensitive content, users can review which files should remain on their machine before Hybrid Compute delegates work across models1

1

.

The architecture functions like a dispatcher, with a cloud-based frontier model breaking tasks into subtasks and routing each to the appropriate location4

4

. Web research, long-horizon planning, and heavy reasoning run in the cloud, while anything touching private files or local data gets delegated to a subagent running on the Mac itself4

4

. Staff detailed the handoff process: "The cloud orchestration will break down the task based on the prompt and figure out how to route it to different subagents... it's going to delegate that down to a sub-agent running on your Mac, and then that portion of the task is run entirely local. None of those tokens go to the cloud"4

4

.

Local Model Options and Cost Efficiency Through Token Savings

Users can select from three local LLMs for privacy-focused processing: Google's Gemma E4B, Alibaba's Qwen 3.6 35B model, and a Perplexity post-trained version of Qwen 3.6 35B, which serves as the company's recommended option1

1

5

5

. Installing a local model requires no terminal access, as Perplexity's app handles the entire setup process1

1

. The company plans to offer additional local models in future updates1

1

.

Hybrid Compute delivers significant cost advantages by eliminating token fees for local inference. Users are not charged for tokens generated by local models on their personal machines, paying only for cloud orchestration and delegation2

2

4

4

. Staff explained the economics: "You're paying for the electricity, you're paying for the hardware, so we're not charging you for that. The only thing the credits are used for is the orchestration and the delegation"4

4

. During processing, the app displays real-time visualization of local CPU, GPU, and memory usage alongside a sidebar tracking token consumption1

1

.

Real-World Use Cases from Legal to Private Equity

Perplexity demonstrated Hybrid Compute through scenarios designed for professionals handling confidential data off the cloud. In one demonstration, a lawyer updated a draft brief against privileged case files stored locally while a cloud agent simultaneously pulled public case law from the web, sending only anonymized legal questions4

4

. "At no point did their privileged information get shared to the cloud. It never left the Mac," Staff noted4

4

.

A second demonstration showed a private equity associate's agent reworking a financial model against confidential management projections while benchmarking the deal against public comparables to produce a fifth iteration of an investment committee deck4

4

. The task ran approximately 40 minutes in the background with no human input, replacing hours of manual work stitching between local spreadsheets and cloud research4

4

. A third demo emphasized cross-device continuity, with a pottery shop founder initiating a marketing analysis from her iPhone that triggered her Mac to process customer interviews and revenue data locally while combining cloud research on competitors' pricing4

4

.

Performance Trade-offs and User Control Over Privacy Scale

Staff acknowledged that fully cloud-based outputs typically deliver superior quality for raw artifact creation: "The short answer is that a fully frontier output is going to almost always be better in terms of raw artifact creation. It's more expensive, it's more capable"1

1

5

5

. However, he emphasized that some users prioritize data privacy and cost efficiency over maximum model performance, particularly for work that doesn't require the most powerful cloud-based frontier models1

1

.

"I think it's got to be a sliding scale, and we want the user to have control over where they are on that sliding scale for the particular type of work they want to do," Staff explained1

1

. He added that Perplexity will suggest optimal configurations based on available context, but ultimately users decide which solution fits their needs1

1

. After the system generates an output, users can write follow-up instructions as usual or queue tasks from an iPhone1

1

.

System Requirements and Questions About Chinese-Developed Models

Hybrid Compute currently runs exclusively on Apple Silicon Macs with macOS 15 or later2

2

. Perplexity recommends machines with at least 32GB of unified memory for optimal performance, though 24GB serves as the minimum requirement3

3

5

5

. Macs with 8GB and 16GB RAM cannot run the feature3

3

.

The inclusion of Qwen, a Chinese-developed model from Alibaba, as the recommended local option has raised questions about enterprise adoption. When asked whether enterprise customers or government clients expressed concerns about giving a Chinese-developed model access to their machines, Staff argued that local inference neutralizes such risks, though the article cuts off before completing his response4

4

. The feature represents Perplexity's latest advancement in agentic AI following releases of Personal Computer in April and Portable Computer in late May2

2

. Apple specifically cited Perplexity Personal Computer as a productivity use case for the new M6 Mac mini3

3

.