Apple Foundation Models Workbench
A workbench for evaluating Apple’s OS 27 Foundation Models using repeatable text and image inputs across on-device and Private Cloud Compute models.
OS 27 introduces more capable on-device models with text and image understanding, alongside access to larger server-based models through Private Cloud Compute. This project provides a quick way to explore how those models perform for a specific product idea before building the complete feature.
Choose a model, provide text or an image, define the desired JSON structure, and run the same prompt multiple times. The workbench groups matching responses and shows their level of agreement alongside latency, making it easier to refine instructions and understand how consistently a model behaves.
A shared core supports a local HTTP server, a native macOS app, and an iOS app.
The project began as an experiment using outdoor security-camera images. Rather than treating each camera event as a simple motion trigger, I wanted to add meaningful context about what was happening—producing structured information that another system could use for better notifications and automation.