Run compatible GGUF local LLMs directly in Unity Android applications through an on-device native runtime—without an HTTP server, WebSocket server, Python environment, or cloud service.Android Edge AI Runtime for Unity enables Unity Android projects to run compatible GGUF-format local language models directly on the Android device.Unity C# communicates with a native llama.cpp-based runtime through the included native shim. Local inference runs inside the Android application without requiring an HTTP server, WebSocket server, Python environment, or cloud service.The package includes:• A minimal Android chat sample scene• Unity C# runtime components• Android arm64-v8a native libraries• UTF-8-safe multilingual text handling• Automatic chat-template application for compatible models• Single-flight generation control to prevent overlapping inference requests• Runtime and model configuration options• Setup and model-placement documentation• Third-party license noticesGGUF model files are not included. Developers select and place their own compatible GGUF model on the Android device.This package is intended for developers building:• Offline-capable AI applications• Interactive characters and virtual agents• Local AI prototypes• Robotics interfaces• Edge AI demonstrations and experiments• Unity applications that require on-device language-model inferenceRequirements and limitations:• Android only• arm64-v8a only• IL2CPP scripting backend required• CPU inference backend• Inference must be tested on an Android device• Unity Editor inference is not supported• Performance and memory usage depend on the Android device, model architecture, model size, quantization, context size, and runtime settings• Compatibility with every GGUF model is not guaranteedThird-party software:This package uses llama.cpp and ggml under the MIT License, and LLVM libc++ under the Apache License 2.0 with LLVM Exceptions.See the included THIRD_PARTY_NOTICES.md and license files for complete details.Key features:• Direct Unity C# to Android native runtime integration• GGUF local language-model inference through llama.cpp• Included native bridge/shim for Unity integration• On-device and offline-capable inference• No HTTP server required• No WebSocket server required• No Python environment required• No cloud service required for local inference• UTF-8-safe multilingual input and output handling• Automatic chat-template application for compatible models• Single-flight generation control• Configurable context size, generation length, sampling, stop sequences, and runtime parameters• Minimal Android chat sample scene• Setup, Android build, and model-placement documentationRuntime architecture:Unity C#→ P/Invoke→ Android native shim→ llama.cpp→ GGUF local language modelIncluded Android native libraries:• libc++_shared.so• libggml-base.so• libggml-cpu.so• libggml.so• libllama.so• libllama_unity_shim.so• libmtmd.soSupported environment:• Unity 2022.3 LTS• Android• arm64-v8a• IL2CPP scripting backend• CPU inference backendUnity Editor execution:The runtime does not execute local LLM inference inside the Unity Editor. Build and run the project on a supported Android device.Model placement:Place a compatible GGUF model on the Android device at:Android/data/{your.package.name}/files/models/model.ggufThe actual package name must match the Android application identifier configured in Unity.Important notes:• GGUF model files are not included• This package provides the local inference runtime and Unity integration layer• Model compatibility is dependent on the llama.cpp runtime included with the package• Inference speed varies significantly between Android devices• Large models or context sizes may exceed the available device memoryAI tools were used as development assistants for code review, documentation drafting, English localization, naming consistency checks, and package structure organization.The runtime architecture, native integration, implementation decisions, Android device testing, debugging, and final verification were performed by the developer.All AI-assisted suggestions were manually reviewed, tested, and revised before inclusion in the package.
