flutter_device_ai 0.2.0

On-device generative AI for Flutter: Apple Foundation Models, Gemini Nano and Chrome built-in AI with streaming, tool calling and JSON-schema output.

0.2.0

• Built-in local fallback model. FlutterDeviceAi.withLocalModel() / LocalLlmFallbackAdapter run a real open GGUF (or .litertlm) model with llama.cpp through llamadart when the OS has no model. prepare() downloads it. OnDeviceAiFallbackAdapter.prepare() added.
• Breaking: EdgeAiFallbackAdapter is now CallbackFallbackAdapter; OnDeviceAiBackend.fallbackEdgeAi is now OnDeviceAiBackend.fallback. The plugin no longer refers to flutter_edge_ai.
• Breaking: minimums raised to Flutter 3.38 / Dart 3.10.7, iOS 16.4, macOS 14 (llama.cpp). Apps need Swift Package Manager enabled on iOS and macOS (the llama.cpp companion is SwiftPM-only), so CocoaPods-only apps must migrate. The plugin manifest declares INTERNET on Android.
• Local fallback quality: conversations use the model's own chat template through structured FallbackMessages (OnDeviceAiFallbackAdapter.generateChat / generateChatStream, default renders a transcript), and JSON-schema answers are grammar-constrained. The emulated tool prompt now includes a worked example.
• Local fallback downloads and storage: resumable downloads with retries, optional sha256, allowDownload hook (for example Wi-Fi only), cancelPrepare(), typed errors for network and disk failures, status stream (idle, downloading, loading, ready, unloaded), isDownloaded(), deleteModel(), presets and recommendedFor(ramGb).
• Local fallback memory: idleUnloadAfter and unload() release the model from RAM; the next request reloads it from disk without downloading.
• Failover: a session on the OS model that reports unavailable mid-use moves to the fallback and answers the turn there.
• Android: ML Kit GenAiException codes map to CANCELLED, SESSION_BUSY, UNAVAILABLE, UNSUPPORTED.
• Generated UI: table rows from small models are padded or trimmed to the header width, fixing a DataTable assertion.
• Example app: local model picker (presets or any GGUF), live download/load status, cancel and delete buttons, 5-minute idle unload, the real fallback instead of the echo stub, macOS network entitlement, and an integration test that downloads and runs a real local model.
• Docs: README, llms.txt and the routing diagram describe the local fallback.
• Apple: availability now probes the model once with a tiny generation, so a missing or unusable model (for example the iOS simulator on a Mac without Apple Intelligence, ModelManager error 1026) is reported as unavailable and routes to the fallback adapter. ModelManager errors map to UNAVAILABLE.

0.1.1

• Swift Package Manager support on iOS and macOS (CocoaPods still works). Sources now live in ios|macos/flutter_device_ai/Sources/flutter_device_ai.
• Shorter pubspec description for pub.dev search.

0.1.0

Package renamed to flutter_device_ai (pub.dev rejected flutter_on_device_ai as too similar to an existing package). Import package:flutter_device_ai/flutter_device_ai.dart; the facade class is FlutterDeviceAi.

First public release.

Engines

• Apple Foundation Models on iOS and macOS 26+: text, streaming, cancellation, native tool calling, schema-constrained output.
• Gemini Nano on Android through the ML Kit GenAI Prompt API: text, streaming, cancellation, image input, exact token counts, model download with progress, Play Store redirect for Google AICore.
• Chrome / Edge built-in AI on the web through the Prompt API: text, streaming, cancellation, native responseConstraint schemas, exact token counts, download with progress and user-gesture detection.
• Fallback adapters for devices without an OS model: OnDeviceAiFallbackAdapter and EdgeAiFallbackAdapter (later renamed CallbackFallbackAdapter). Fallback works for sessions, streaming, tools and schemas, not only one-shot calls.
• Windows and Linux report unavailable and route to the fallback adapter.

API

• FlutterDeviceAi: generateText, generateTextStream, generateTextSimple, createSession, availability, isAvailable, capabilities, getPlatformInfo, prepare / ensureReady, downloadModel, openAICorePlayStore, initialize, registerTools.
• OnDeviceAiSession: concurrent isolated sessions, multi-part turns, ask / askStream, countTokens, real stopGeneration, busy detection.
• Emulated tool calling for backends without native tools, with argument validation and a bounded round-trip loop (also while streaming).
• Structured output with a built-in JSON Schema validator, lenient JSON extraction (fences, prose, truncation) and automatic validate-and-repair.
• OnDeviceAiCapabilities reports each feature as native, emulated or unavailable.
• OnDeviceAiAvailability carries a status, a machine-readable reason and a UI-ready message.
• Typed exception hierarchy under OnDeviceAiException.
• Generative UI: OnDeviceUiGenerator, GenUiModuleSpec with eight validated block types (stat, progress, checklist, list, steps, table, callout, note), localisation, streaming progress and tool support.
• package:flutter_on_device_ai/testing.dart with FakeOnDeviceAiPlatform.

Tooling

• 60+ unit tests, a browser test suite for the web backend, and a device integration test that exercises the real model and skips itself where none exists.
• CI builds Android, iOS, macOS, web, Linux and Windows example apps.
• llms.txt and AGENTS.md for AI assistants.

Not yet

• Windows AI Foundry (Phi Silica) backend.
• Image input on Apple platforms and on the web.
• Exact token counts on Apple platforms.