Archive · 1 Sep 2026
Perplexity Launches Hybrid Compute For Sensitive Local-Cloud AI Tasks
Perplexity created a tool called Hybrid Compute for Mac computers.
It lets some AI work happen online and some happen directly on the user’s computer.
Keeping certain work on the computer can help protect private information.
It may also reduce the amount of work sent to more expensive online AI models.
Users can check which files will be used before starting.
They can also choose which AI models handle the work.
Local models currently include Gemma E4B and versions of Qwen 3.6.
Work done by a local model does not count toward token charges.
The tool currently works only on certain Macs and is limited to Pro and Max subscribers.
Perplexity’s Hybrid Compute splits tasks between cloud models and local models on a user’s Mac.
The tool is designed to keep sensitive information on the user’s computer while potentially reducing cloud inference costs.
Users can review selected files and choose which models process different parts of a task.
Supported local models include Gemma E4B and two Qwen 3.6 versions with 35 billion parameters.
Hybrid Compute currently supports Apple Silicon Macs running macOS 15 and requires a Pro or Max subscription.
- Who
- Perplexity announced Hybrid Compute for users of its AI tools.
- What
- A feature that combines cloud-based frontier models with local language models to divide task processing.
- When
- The article does not specify the announcement date.
- Where
- On Apple Silicon Macs running macOS 15, with local processing taking place on the user’s computer.
- Why
- To help keep sensitive data local and potentially reduce inference costs.
- Availability
- Currently available only on Apple Silicon Macs running macOS 15.
- Subscription
- Available only to Perplexity Pro and Max subscribers.
- Recommended hardware
- Perplexity recommends at least 32GB of unified memory.
- Local models
- Current options include Gemma E4B and two versions of Qwen’s 35-billion-parameter 3.6 model.
- Privacy feature
- Users can keep sensitive information on their local computer while cloud and local models share the work.
- Cost feature
- Tokens generated by a local model on the user’s machine are not charged.
- Task controls
- Users can review selected files, choose models, monitor hardware usage, and view token consumption.