firebase-ios-sdk/GeminiLanguageModel/FoundationModelsFake

FoundationModelsFake

Important

FoundationModelsFake is a fake implementation of Apple's Foundation Models framework. It supports development and testing where the real framework isn't available, such as on Linux. Its behavior can differ from the real framework. Don't ship it in production apps, and test your app against the official Foundation Models framework before release.

This package lets GeminiLanguageModel build and run its tests on Linux. It mirrors the public API of Apple's framework in the iOS 27.2 and macOS 27.2 SDKs (Xcode 27.2) closely enough that the Gemini sources compile against either one.

Using the fake

GeminiLanguageModel/Package.swift depends on this package through its FoundationModelsFake trait. When the trait is enabled, SwiftPM also defines a FoundationModelsFake compilation condition, which the Gemini sources check:

#if compiler(>=6.4) && (FoundationModelsFake || canImport(FoundationModels))
  #if FoundationModelsFake
    import FoundationModelsFake
  #else
    import FoundationModels
  #endif
#endif

Where	How to enable the trait
Linux	It's on by default.
macOS command line	swift build --traits FoundationModelsFake or swift test --traits FoundationModelsFake
Xcode project or workspace	Enable the trait on the package dependency in the Package Dependencies view.
Cross-compiling for Linux from macOS	Pass --traits FoundationModelsFake, because the default is chosen by the machine that runs SwiftPM.

With the trait off, SwiftPM neither fetches nor builds this package.

To run the fake's own tests, run swift test in this directory. These are unit tests; the fake's full behavior-comparison suite is maintained separately.

Differences from Apple's framework

• Removed types: PrivateCloudComputeLanguageModel, OCRTool, BarcodeReaderTool, the obsolete SystemLanguageModel.Adapter API, and the feedback API (LanguageModelFeedback and LanguageModelSession.logFeedbackAttachment).
• Deprecated APIs are omitted, such as LanguageModelSession.GenerationError, Transcript.StructuredSegment.source, and GenerationOptions.sampling.
• SystemLanguageModel is always unavailable. Its availability is .unavailable(.deviceNotEligible), and responding throws SystemLanguageModel.Error.assetsUnavailable. Pass a model, such as GeminiLanguageModel, when you create a session.
• Swift versions: below Swift 6.4, the fake compiles to an empty module, like Xcode 26 SDKs, which lack the iOS 27 LanguageModel API. The iOS 27.2 API, such as Transcript.DataEntry, needs Swift 6.4.2, like Xcode 27.2.
• Availability: public declarations carry the same @available attributes as Apple's, so code that uses the fake on Apple platforms is checked the same way. Availability checking is disabled inside the fake itself.
• Images on Linux: image attachments are backed by URLs, because Linux has no CoreGraphics or ImageIO, and their embedded EXIF orientation isn't read. Linux also gets small shims for UTType and CGImagePropertyOrientation, which cover only the types the fake needs.
• Image attachments compare by identity, not by pixels. A cgImage that can't be loaded, such as one from a remote URL, is a transparent 1×1 placeholder.
• Transcript JSON uses the same format as Apple's, on purpose, so that transcripts can be exchanged with the real framework. Images backed by URLs are written as their URL (file: URLs are embedded) instead of being re-encoded, and remote URLs are never fetched. Decoding rejects file: image URLs.
• Decoded structured content isn't == to the content it was encoded from, because decoding loses the explicit key order. Their JSON is the same.
• Messages: error and description strings are the fake's own, so don't match on their text.