flutter_gemma 0.11.13

The plugin allows running the Gemma AI model locally on a device from a Flutter application. Includes support for Gemma 3 Nano models with optimized MediaPipe GenAI v0.10.24.

0.11.13

• ✅ iOS Embeddings Fix: XNNPACK + SentencePiece integration for better results on iOS
• 🌐 Web CDN: Modules available via jsDelivr (@0.11.13/web/*.js)

0.11.12

• 🌐 Web VectorStore: Full RAG support on web with SQLite WASM
  • Uses wa-sqlite with OPFS storage (10x faster than IndexedDB)
  • Cross-platform parity: Android, iOS, and Web now all support VectorStore
• 🐛 Android 16KB Page Size Fix: Updated tasks-vision-image-generator to 0.10.26.1 for Android 15+ compatibility
• ⬆️ Kotlin Update: Upgraded to Kotlin 2.1.0 in example app

0.11.11

• 🐛 Mobile Build Fix: Fixed compilation errors on iOS/Android platforms

0.11.10

• 💾 Web Persistent Caching: Models now persist across browser restarts using Cache API
• ⚠️ BREAKING CHANGE: Explicit initialization now required
  • ACTION REQUIRED: Add await FlutterGemma.initialize() in main() before using the plugin

0.11.9

• 🌐 Web Embedding Support: Added support for embedding generation on web platform
• 🐛 Web Example App: Fixed bugs in example app on web platform

0.11.8

• 🤖 CI/CD Automation: Added GitHub Actions workflows for automated testing and release builds
• ☕ Support the Project: Added Ko-fi donation button for community support

0.11.7

• 🚀 VectorStore Optimization ⚠️ BREAKING (RAG only):
  • 71% smaller storage: Binary BLOB format instead of JSON (3 KB vs 10.5 KB per 768D embedding)
  • 6.7x faster reads: ~75 μs vs ~500 μs per document search
  • 3.3x faster writes: ~45 μs vs ~150 μs per document insertion
  • Dynamic dimensions: Auto-detect any embedding size (256D, 384D, 512D, 768D, 1024D, 1536D, 3072D, 4096D+)
  • iOS implementation: Full VectorStore support on iOS (was stubs only)
  • Cross-platform parity: Identical behavior on Android and iOS
  • ⚠️ ACTION REQUIRED: Existing vector databases will be recreated on upgrade (re-index required)

0.11.6

• 🐛 iOS Simulator Fix: Fixed "Filename cannot contain path separators" crash on iOS Simulator (#127)
• 🔧 Download Service Refactoring: Unified download implementation, removed legacy code (~100 lines)
• 🚫 Download Cancellation: Added CancelToken pattern (Dio-style) for cancelling downloads (NON-BREAKING)
• 🧹 API Cleanup ⚠️ BREAKING: Removed unused canResume(), resume(), cancel() methods from DownloadService interface

0.11.5

• 🐛 Fixes: Some fixes for new Modern API
• ⚠️ Deprecated: Marked legacy asset/file management methods as deprecated with migration hints
• 📚 Documentation: Updated README with Modern API examples and complete parameter documentation

0.11.4

• New: Fluent builder API with FlutterGemma.installModel().fromNetwork/fromAsset/fromBundled/fromFile()
• Architecture: Sealed classes for type-safe sources, dependency injection via ServiceRegistry, platform-specific handlers
• Bundled Models: Support for including small models in production builds via native bundles (iOS: Bundle.main, Android: assets)
• Backward Compatible: Legacy API (modelManager.downloadModelWithProgress()) still works as facade

0.11.3

• 🌐 Web Multimodal Support: Added full multimodal image processing support for web platform
• 📚 MediaPipe 0.10.25: Updated to MediaPipe GenAI v0.10.25 for web compatibility
• 📦 LiterTLM Format Support: Added support for .litertlm model files optimized for web platform

0.11.2

• 🛡️ Fixed: Updated ProGuard rules for Android release build compatibility

0.11.1

• 🐛 Fixed: Export missing ModelType and other public API types to resolve import issues

0.11.0

• 🚀 Embedding Models Support: Added full support for text embedding models
• 🔧 Unified Model System: All models (inference and embedding) now use the same download and management pipeline
• 📝 ModelSpec Architecture: Introduced InferenceModelSpec and EmbeddingModelSpec for better model organization
• 🛡️ Smart Cleanup System: Added automatic cleanup of orphaned files with resume detection capabilities
• 🔄 Model Replace Policies: Separate policies for model downloading replace and keep
• 📱 Example App Integration: Added embedding models download screen and embeddings generation demo screen

0.10.6

• 🔧 Model Replace Policy: Added configurable model replacement system with keep/replace options and ensureModelReady() method
• 📥 Enhanced Downloads: Added HuggingFace CDN ETag issue handler with smart retry logic and exponential backoff
• 🔄 Download Reliability: New HuggingFaceDownloader wrapper to handle CDN server inconsistencies and resume failures
• 📁 ModelFileType System: Introduced distinction between .task files (MediaPipe-handled) and .bin/.tflite files (manual formatting)
• 🔐 Android Security: Added network security configuration for HuggingFace CDN access with proper permissions
• 🐛 Download Fixes: Fixed Android download timeouts, stream management, and ETag mismatch issues
• 🖼️ Image Corruption Fix: Added comprehensive image processing system to prevent AI model corruption on Android
• 🔄 Example App: Added sync/async response method selection in chat interface

0.10.5

• 🛑 Stop Generation: Added Android support for stopping text generation mid-process with session.cancelRequestGenerationAsync() (#89, #19, #34)
• 🐛 Screen Close Fix: Fixed crash when closing screen during active generation by implementing proper StreamSubscription cleanup (#89)
• 🐛 Model Loading Fix: Fixed model install check with partial downloads by adding orphaned files cleanup (#84)
• 🗑️ File Management: Added automatic cleanup of corrupted/incomplete model files with atomic SharedPrefs updates
• 📱 iOS Requirements: Updated deployment target to 16.0 for MediaPipe GenAI compatibility
• 🔧 Error Handling: Improved error recovery with automatic file cleanup on failed downloads
• 📚 Documentation: Updated model capabilities table and comprehensive usage examples

0.10.4

• 📚 Documentation: Updated README with comprehensive model information and usage examples

0.10.3

• 📥 Background Downloads: Added background download support for model files

0.10.2

• 🚀 New Models: Added support for 4 new compact models:
  • Gemma 3 270M - Ultra-compact text-only model (0.3GB)
  • TinyLlama 1.1B - Lightweight chat model (1.2GB)
  • Hammer 2.1 0.5B - Action model with strong function calling (0.5GB)
  • Llama 3.2 1B - Instruction-tuned model (1.1GB)
• ⚙️ Backend Switching: Added individual CPU/GPU backend switcher for each model in the example app
• 🔍 Advanced Model Filtering: Added expandable filter section with feature-based filtering:
  • Filter by Multimodal support (vision capabilities)
  • Filter by Function Calling support
  • Filter by Thinking Mode support
  • Animated expandable UI with clear visual indicators
• 📊 Model Sorting: Added comprehensive sorting options:
  • Sort alphabetically (A-Z)
  • Sort by model size (smallest to largest)
  • Default order (Gemma models prioritized)
• 📏 Improved Model Display: Separated size information from model names for cleaner presentation
• 🌐 Full English Localization: Converted all UI text from Russian to English
• 📊 Results Counter: Added dynamic counter showing filtered results ("Showing X models")
• 🎨 Enhanced Filter UI: FilterChip components with color-coded selections matching feature badges
• 🎯 Model Organization: Reorganized model list with Gemma models prioritized at the top
• 🛠️ Function Calling: Enhanced function calling support with Hammer 2.1 action model
• 📱 UI Improvements: New card-based design with individual backend controls per model
• ✅ Model Capabilities: Fixed and verified multimodal support flags for all models
• 📚 Documentation: Updated README and model feature support table with new models

0.10.1

• 🧠 Thinking Mode: Added thinking mode support for DeepSeek models with persistent thinking bubbles
• 🔧 Function Call Fixes: Fixed function calls detection in the middle of responses
• 💬 UI Improvements: Fixed loading indicator display after function calls
• 🔄 JSON Response Handling: Enhanced handling of JSON responses after function execution
• 📚 Documentation: Updated README with latest API changes and improved examples
• 🎨 Code Quality: Removed Russian comments and improved code consistency

0.10.0

• ✨ Function Calling: Added support for function calling, allowing models to interact with external tools.
• 🔧 Improved Chat API: Enhanced the chat API to support function calls and tool responses.

0.9.0

• 🖼️ MULTIMODAL SUPPORT: Added full support for text + image input with Gemma 3 Nano vision models
• 🎯 Enhanced Message API: New Message class with support for text, image, and multimodal content
  • Message.text() - for text-only messages
  • Message.withImage() - for text + image messages
  • Message.imageOnly() - for image-only messages
  • message.hasImage - to check if message contains image
• 📱 Vision Models: Full support for Gemma 3n E2B and E4B models with image understanding
• 🌐 Web Platform: Added graceful degradation with debug warnings for unsupported features

0.8.6

• 🚀 GEMMA 3 NANO SUPPORT: Added full support for Gemma 3 Nano models
• ⚡ Optimized session parameters for Gemma 3n models (temperature: 0.6, topK: 40, topP: 0.9)
• 🛡️ Added automatic fallback session creation for input_pos != nullptr errors
• 🎯 Added Gemma 3n model detection and compatibility handling
• 💪 Enhanced error handling for TensorFlow Lite model initialization
• 🔧 Fixed iOS session initialization with proper input position handling
• 📱 Improved mobile inference model with optimized parameters

0.8.5

• Upgraded Mediapipe to 0.10.24 for iOS and Android
• Added support of Gemma3, Phi-4 and DeepSeek models for iOS

0.8.4

• iOS LoRA support added
• iOS topP support added

0.8.3

• Readme updated

0.8.2

• Readme updated

0.8.1

• Migrate to js-interop
• Add web platform support in pubspec.yaml

0.8.0

• Upgraded Mediapipe to 0.10.22 for Android and Web
• Upgraded Mediapipe to 0.10.21 for iOS
• Added opportunity to set topP and preferredBackend for inference
• Added support of Gemma3, Phi-4 and DeepSeek models for Android and Web

0.7.0

• Added Chat functionality for instruction tuned model
• Added sizeInTokens method for analysis of the size of the input prompt

0.6.0

• Added opportunity to manage inference session

0.5.1

• Fixed crash on generation for Android

0.5.0

• IMPORTANT: Breaking changes in the API
• FlutterGemma instance was replaced with ModelManager and InferenceModel
• ModelManager to manage models and LoRA weights
• InferenceModel to manage inference
• Added opportunity to set model and LoRA weights paths manually
• Added opportunity to delete model and LoRA weights
• Added opportunity to load LoRA weights from assets and network

0.4.6

• Added close method

0.4.5

• Small fixes for Android

0.4.4

• Small fixes for iOS

0.4.3

• Upgraded Mediapipe to 0.10.20
• Updated LoRA support

0.4.2

• Added error handling
• Updated example for error handling
• Upgraded Mediapipe to 0.10.18 for iOS
• Fixed ios issue with model freezing

0.4.1

• Fixed ios issue

0.4.0

• Upgraded Mediapipe to 0.10.16
• Added LoRA support
• Fixed some issues

0.3.1

• Updated example and readme

0.3.0

• Added support for loading models from assets and network.
• Added progress updates for model loading.

0.2.4

• Fixed Mediapipe ios version

0.2.3

• Updated Mediapipe ios version

0.2.2

• Added opportunity to configure folder (Android only)
• Fixed android release issue
• Updated Mediapipe for Android
• Updated readme

0.2.1

• Updated chat functionality for instruction tuned model
• Updated readme

0.2.0

• Added chat functionality for instruction tuned model
• Updated example

0.1.4

• Updated readme for GPU models on Android devices
• Updated example

0.1.3

• Updated readme

0.1.2

• Updated example

0.1.1

• Updated example

0.1.0

• Added web support
• Added opportunity to set randomSeed, topK and temperature

0.0.4

• Updated example
• Updated readme

0.0.3

• Added getResponseAsync method

0.0.2

• Added description in Readme.md
• Added opportunity to setup a model before initiation

0.0.1

• Initial release