apple_foundation_models 0.1.0
apple_foundation_models: ^0.1.0 copied to clipboard
Run Apple's on-device language model from Flutter. Streaming, tool calling, and schema-constrained structured output — private, offline, and free.
0.1.0 #
Initial release.
-
Generation guides:
Schema.integer(min:, max:),Schema.number(min:, max:),Schema.string(pattern:),Schema.array(exactItems:). Bounds constrain generation itself, so a 1-to-5 rating cannot come back as 47. -
respondInto/streamIntodecode straight into your own types. -
deltas()turns cumulative snapshots into increments. -
ModelUseCase.contentTaggingselects Apple's tagging-tuned model. -
AppleFoundationModels.supportedLanguages()andavailabilityChanges, the latter firing when the model finishes downloading mid-session. -
AppleFoundationModels.availability()with typed reasons —deviceNotEligible,appleIntelligenceNotEnabled,modelNotReady,osTooOld,unsupportedPlatform— each carrying an explanation and, where one exists, a remedy. -
LanguageModelSessionwith instructions, transcript,prewarm, andisResponding. -
Four response modes:
respond,respondAs(schema-constrained),stream, andstreamAs(structured output arriving progressively). -
A Dart schema DSL — objects, arrays, enums, optional fields — rebuilt natively as a
DynamicGenerationSchema, so typed output needs no codegen. -
Tool calling: the model invokes a Dart function mid-generation and waits for the result.
-
GenerationOptionswith temperature, token cap, and greedy / top-k / top-p sampling. -
Every framework error mapped to its own exception type.
-
Compiles and runs on every platform. Non-Apple targets report
unsupportedPlatforminstead of failing the build, and an Xcode without the iOS 26 SDK still compiles.