ExecuServe

Alpha Romer Coma

About this app

ExecuServe turns your Android phone into a local AI server. It keeps compiled ExecuTorch models loaded on the phone and answers the OpenAI and Anthropic APIs over HTTP, so apps on the phone, or on your network if you allow it, can use them like any hosted model.

SERVE
• OpenAI Chat Completions and Responses, and Anthropic Messages, with streaming and tool calls
• Several models kept ready at once, sharing one queue
• Keeps serving in the background, with an ongoing notification and a Stop action
• This phone only by default; your network only when you choose it, always with API keys

CHAT
• Talk to any installed model from the Chat tab, through the same API your apps use
• Turn thinking on or off for models that support it
• Prefill and decode speeds on every reply

CONNECT
• Base URLs, model IDs and keys to copy, or to scan as QR codes
• A browser chat built in, so a laptop or tablet can talk to the phone's models with nothing installed

KNOW WHAT HAPPENED
• Every request with its phases, speeds and device state, kept across restarts
• Like-for-like comparison of models, a benchmark, and CSV export

MODELS
• Download ready-made ExecuTorch exports from Qwen, Meta, Liquid AI and others, published by Experimental Machines on Hugging Face
• Or copy your own .pte model and tokenizer to the phone from a computer

PRIVATE BY DESIGN
• Replies are generated on the phone. There is no account, no analytics and no ExecuServe server.
• Over the internet, the app contacts only Hugging Face: for the catalog, model downloads and publishers' pictures.
• Every reply has "Report this reply"; you see the whole report before anything is sent.

AI models can make mistakes and may produce inaccurate or offensive content. Replies come from the model you choose, not from ExecuServe.

Requires Android 12 or newer on a 64-bit Arm phone, and free memory for the models you keep loaded: about each model's file size.

ExecuServe is open source under the Apache 2.0 licence: github.com/ExperimentalMachines/execuserve

An independent project by Experimental Machines. Not affiliated with, endorsed by or sponsored by the PyTorch Foundation or Meta

What’s new

• GPU models: the catalog now offers Vulkan (GPU) builds on phones whose GPU can run them, next to the CPU builds. If your GPU can't run one, ExecuServe remembers and offers the CPU build instead.
• /v1/models now says whether each model runs on the CPU or the GPU
• Updated to the ExecuTorch 1.5.1 runtime; models exported for 1.4 keep working.