Availability

How to interpret getAvailability() statuses and platform-specific runtime behavior. Related guides: Setup, Android fallback model, Chat, Images, Events.

Availability

getAvailability() returns a semantic status value:

Status	Meaning
available	The text model is ready for generation.
device-not-eligible	The device or OS version does not support the on-device text model. iOS reports this precisely; Android's FeatureStatus.UNAVAILABLE does not expose a reason and maps to unavailable.
not-enabled	The device supports on-device AI but the user has not enabled it (e.g. Apple Intelligence is off). Primarily reported on iOS.
downloadable	The model can be downloaded (Android).
downloading	A model download is in progress (Android).
not-ready	The model exists but is still initializing.
unavailable	The model is unavailable for another reason.

Subscribe to availabilityChange (or the deprecated systemAvailabilityChange alias) to receive updates while listeners are registered. On Android the plugin polls while listeners are active.

The deprecated systemAvailability() and systemAvailabilityChange fold detailed statuses into the original four-value contract: available, unavailable, notready, and downloadable. downloading and not-ready map to notready; device-not-eligible, not-enabled, and unavailable map to unavailable.

Platform Behavior

iOS

• Text LLM requires iOS 26 and Apple Intelligence. Below iOS 26, availability is reported as device-not-eligible. Only select iPhones (iPhone 15 Pro or later) and iPads support Apple Intelligence. More information here.
• Chats use the native Foundation Models transcript. Conversation state lives in LanguageModelSession. Before generation, the plugin also applies a conservative character budget against the model's native contextSize, including instructions, the current prompt, and output headroom. When any limit is exceeded, it drops the oldest complete prompt/response turns, preserves instructions, and recreates the session.
• warmup({ chatId, promptPrefix }) prewarms a specific chat created with createChat().
• cancelGeneration() cancels the in-flight Task for the chat. The streamText() / generateText() promise rejects with LOCAL_LLM_GENERATION_CANCELLED; any text already streamed via textChunk remains in your UI.
• Image analysis and image generation details are in Images.

Android

• Conversation history is structured in memory by the plugin. Each chat stores user/assistant turns and trims by history.maxMessages (default 20) and history.maxCharacters (default 12000). Gemini Nano then uses ML Kit countTokens() to fit the context. LiteRT-LM uses a conservative one-character-per-token heuristic plus prompt and image reserves because version 0.16.1 has no stable token-count API. Both paths drop oldest whole turns; system instructions are stored separately and never trimmed. History does not persist across app restarts.
• warmup() warms the model globally and ignores chatId / promptPrefix.
• cancelGeneration() is best-effort. It cancels the plugin's coroutine; ML Kit may already have emitted partial output before cancellation completes. The promise rejects with LOCAL_LLM_GENERATION_CANCELLED when cancellation is observed.
• Unsupported GenerationOptions values are rejected with LOCAL_LLM_INVALID_OPTIONS, not silently clamped. maxOutputTokens must be within 1..min(device token limit, 4096); when omitted, the plugin default is 256.
• Native model operations run serially. The plugin mutex protects Gemini Nano and LiteRT-LM generation, fallback configuration, warmup, download, and teardown, so concurrent generations in different chats are queued.
• Not all API 29+ devices support Gemini Nano. The device must have a compatible on-device AI stack. More information here.
• Apps may explicitly configure a LiteRT-LM fallback when Gemini Nano is unavailable. Once initialization completes, getAvailability() reports available. See Android fallback model.
• Image analysis backend selection is documented in Images.
• On-device models cannot be used while the app is in the background. Inference requests made while backgrounded will fail.
• AICore enforces per-app inference quotas. Excessive requests can return busy or quota errors from the underlying SDK — consider exponential backoff.

Web (Chrome)

Text availability maps Chrome's available, downloadable, downloading, and unavailable states directly. Missing APIs report unavailable. Availability events reflect changes observed by plugin checks and session creation/download, rather than background polling. Image analysis currently reports unavailable. See Web for setup, supported methods, and generation controls.