Apple Foundation Models (iOS 26)¶

AppleFoundationModel wraps FoundationModels.LanguageModelSession. Zero download, always offline, Apple's ~3B on-device model (2-bit QAT weights, 8-bit KV cache).

let avail = AppleFoundationModel.availability()          // nonisolated static
if avail.available {
    let model = AppleFoundationModel()
    try await model.load(spec, from: nil) { _ in }       // nothing to load; marks loadedModelId
    let agent = Agent(model: model, tools: [CurrentTimeTool(), CalculatorTool()])
} else {
    print(avail.reason)                                  // show it verbatim
}

Availability¶

SystemLanguageModel.default.availability → RuntimeAvailability:

Apple	`reason` the SDK returns
`.available`	`.ok`
`.unavailable(.deviceNotEligible)`	"device not eligible for Apple Intelligence"
`.unavailable(.appleIntelligenceNotEnabled)`	"Apple Intelligence is off — Settings › Apple Intelligence & Siri"
`.unavailable(.modelNotReady)`	"model not ready (still downloading / device busy)"
iOS < 26	"needs iOS 26 or later"
SDK without the framework	"FoundationModels framework not in this SDK"

Context: 4,096 tokens, hard¶

Apple states it three times in the docs: prompts, instructions, tool definitions and their input and output, generable schemas, and all responses share one 4,096-token window per session. The runtime reports Metrics.contextMax = 4096 and — because the framework exposes no tokenizer — estimates prompt tokens at ~4 characters each with Metrics.estimated = true (the app shows ≈). Overflow surfaces as GenerationError.exceededContextWindowSize, mapped to a ModelError; trim agent.messages and retry.

Tools: manual, not native (v1)¶

Apple's native Tool protocol needs compile-time typed @Generable arguments, which does not fit a toolset defined at runtime by JSON schema. So v1 prompts manually: PromptBuilder.toolSystemPrompt(tools:family: .generic, base:) goes into the session instructions, the model answers with a bare JSON object, and the agent's ToolCallParser(family: .generic) lifts the call. Native bridging is noted for v2.

Streaming¶

streamResponse yields snapshots (accumulated text), not deltas. The runtime diffs consecutive snapshots into ModelEvent.text deltas so the transcript view behaves the same for both runtimes. Earlier turns are flattened into the prompt as User: … / Assistant: … lines because a session's transcript is append-only.

Other errors you will see¶

rateLimited (background or aggressive use — foreground interactive use is not limited), guardrailViolation, refusal, concurrentRequests (one in-flight request per session — the agent serialises), unsupportedLanguageOrLocale.

iOS 27 APIs

Apple's docs already show iOS 27 material: custom LanguageModel providers, MLXLanguageModel, ToolCallingMode, ContextOptions, Private Cloud Compute. None of it is usable on an iOS 26 target; loopl keeps two engines behind the one Model protocol instead.