denizaydogan1902/anvilai

Anvil AI is an Android app that runs large language models directly on the phone, with no cloud connection required. It uses llama.cpp for the model engine and taps the phone's GPU through Vulkan to speed up the math, so responses generate on device instead of being sent to a server. Because nothing leaves the phone, there is no data leakage and no API key to manage. The app lets you download curated GGUF model files from a built in model store, or paste a Hugging Face link to fetch a model directly, with resumable downloads if a transfer gets interrupted. It supports several small open models, including Llama 3.2 in 1B and 3B sizes, Gemma 2 2B, Qwen 2.5 1.5B, and Phi-3 Mini, each with a listed minimum RAM requirement so a user can pick a model that fits their device. Chat responses stream token by token, and the interface applies the correct prompt template for whichever model family is loaded. Chats are stored locally in an encrypted database, and the app includes a thermal manager that scales back GPU usage automatically if the phone overheats or the battery runs low, plus a performance overlay showing tokens per second, RAM use, and GPU layer counts. The interface is built with Jetpack Compose in a dark, high contrast theme, and supports multiple separate conversation histories with rename and delete options, along with markdown and syntax highlighted code blocks in responses. Building the app from source requires Android Studio, JDK 17, the Android NDK, and CMake, plus a separate step to pull in the llama.cpp source code, skipping that step still builds the app but inference returns demo responses instead of real output. A phone needs an ARM64 processor, at least 4GB of RAM, and Vulkan 1.0 support, which covers most devices from 2017 onward. The project states it is provided as is for educational and personal use.