flutter_llama 1.2.0

Flutter plugin for running LLM inference with llama.cpp and GGUF models on Android and iOS

Changelog

All notable changes to this project will be documented in this file.

1.2.0 - 2026-10-11

Added

• Agent API (doc/agent.md):
  • chat / chatStream with the model's own template, tool calls constrained by grammar (ToolSpec, ToolChoice), JSON-schema / GBNF output, and reasoning on or off;
  • classify (label log-probabilities or embeddings), countTokens, countChatTokens, embed, modelInfo;
  • canLoad: whether a model fits the device's memory and disk before downloading;
  • LlamaLifecycle: unload on background, idle or memory pressure.
• Open-model catalogue ModelCatalog / OpenModels: Qwen3, Granite 3.3, Phi-4-mini and SmolLM2, Apache-2.0 or MIT only, with licence, sizes and tasks (doc/models.md).
• Desktop. Linux through FFI onto the shared core (src/core). The Windows build files are included.
• Android GPU.
  • The CPU and Vulkan backends are separate modules loaded at run time, so minSdk stays 24.
  • nGpuLayers: -1 is auto: it offloads only where the GPU beats the CPU (Mali and PowerVR stay on the CPU).
  • getModelInfo()['backends'] lists the devices.
• The example app has an agent tour (open models, chat with a tool, classify).

Fixed

• Android release builds: consumer R8 rules keep the JNI result classes. Before, the first getModelInfo crashed with NoSuchMethodError under R8.
• Android, Linux and Windows build from a pub.dev install: CMake fetches llama.cpp at the pinned commit (945501f5e) when the checkout is not next to the package. Android needs CMake 3.22.1.

Changed

• Requires Dart 3.4 / Flutter 3.22.
• The 1.1 API (loadModel, generate, generateStream, PresetModels) is unchanged.

1.1.2 - 2025-01-28

Changed

• Optimized package size by excluding debug symbols and build artifacts
• Added comprehensive .pubignore to reduce package size from 234 MB to 20 MB

1.1.1 - 2025-01-28

Fixed

• Fixed package dependencies: moved http, path, and path_provider from dev_dependencies to dependencies for proper pub.dev publishing
• Resolved package validation errors for service dependencies

Changed

• Updated package structure for better dependency management

1.0.1 - 2025-01-27

Fixed

• Fixed macOS library build artifacts for pub.dev publishing
• Removed unused imports in multimodal implementation
• Removed unnecessary library name declaration
• Improved .pubignore to reduce package size

Changed

• Updated package metadata

1.0.0 - 2025-01-27

Added

• Initial stable release
• Full support for llama.cpp GGUF models on iOS and macOS
• GPU acceleration via Metal
• CPU optimization via Accelerate framework
• Streaming text generation
• Batch processing support
• Multiple model format support
• Comprehensive documentation
• Example application with demos
• Integration tests

0.1.1 - 2025-01-26

Added

• Multimodal support (vision models)
• Enhanced model loading
• Improved error handling

0.1.0 - 2025-01-25

Added

• Initial beta release
• Basic llama.cpp integration
• iOS and macOS platform support
• Model loading and text generation
• Basic configuration options