What is QVAC? Tether's open-source local AI, explained

QVAC is Tether's 100% free, private and accessible AI. It lets a phone, a laptop or a robot run AI models, without sending anything to a cloud service. In other words, QVAC & Tether never see or access your data. Developers build apps with the QVAC SDK, and anyone can install QV.AC, the local AI assistant app. The models and the training data behind both are available to download for everyone.

What is local AI?

Local AI means the AI model, the trained program that produces the answers, runs on the device in front of you. Most AI apps today send every question to a data centre, run the model there and send the answer back, so they stop working without a connection.

Running the model on your own phone or laptop changes four things:

• Free: there is no subscription and no bill per question, because the work happens on hardware you already own.
• Private: what you type and the files you open stay on your device, so no company reads them.
• Works anywhere: the model needs no network, so it answers on a plane or in a village with no signal. This video shows why unstoppable intelligence matters.
• Sovereign: the model is a file on your device, so no company can switch it off or take it away.

Our Local AI 101 series covers which jobs a small local model handles well, and which machine runs what.

Why QVAC exists

QVAC's mission is to make intelligence a universal right, not a subscription. A monthly AI plan is a small expense in some countries and several days of income in others, and most AI works best in a handful of widely spoken languages. A model on a phone people already own costs nothing per answer and can be trained for a language the big services cover poorly.

Paolo Ardoino, CEO of Tether, says "Not your AI, not your intelligence": an AI running on someone else's server answers to whoever owns that server.

Tether's independent stack

QVAC is the top layer of what Tether calls its independent stack: a set of open technologies that people can run themselves. In his talk in Lugano in October 2026, Paolo walks through the layers QVAC builds on:

• Energy: solar charging kiosks that bring electricity to villages in Africa, used by about 2 million people.
• Communication: Holepunch, a peer-to-peer network that connects devices directly with no servers, and Keet, the messaging app built on it.
• Money: WDK, the Wallet Development Kit, for building digital wallets that only their owner can control, whether a person or an AI agent (a program that carries out tasks on its own).
• Intelligence: QVAC, which downloads models directly from other devices over Holepunch and lets your own devices share computing power. An agent built on QVAC can also keep its own WDK wallet.

What QVAC includes

QVAC has five building blocks. You only need the one that matches what you want to do.

QVAC SDK, for developers

The QVAC SDK is the building block for any app that uses local AI: a Mac or Windows app, an iPhone or Android app, or a server. It is free and open source (Apache 2.0) and installs with one command, in JavaScript or Python. In JavaScript, the same code runs on iOS, Android, macOS, Windows and Linux. One API, a single set of commands, covers text chat, speech to text, text to speech, translation, image and video generation, image understanding, text extraction from images, search over your own documents, and fine-tuning, which means teaching a model new behaviour from your own examples.

To start, read the docs, follow the quick-start guide or pick one of the recipes, small example apps with their code.

QV.AC, the local AI assistant app

QV.AC is a personal AI assistant that runs on your phone and computer, free and with no account. Each of its agents is an assistant set up for one job, and you can connect them to your calendar, email or files. You can also pair your devices: ask a question on your phone and a larger model answers from your computer, over an encrypted peer-to-peer link. It is available on iOS, Android, macOS, Windows and Linux. At the time of writing, in October 2026, QV.AC is in alpha, an early public version.

Psy models

The Psy models are QVAC's own AI models, built small enough to run on a phone or a laptop. Each one specialises in a single job:

• MedPsy answers medical questions offline. Its smaller version, with 1.7 billion parameters (the numbers a model learns in training), scores higher on average across seven medical benchmarks than Google's MedGemma 4B, a model more than twice its size. It gives medical information, not a diagnosis, and does not replace a doctor. Meet MedPsy tells the full story.
• TranslatePsy translates between English and 19 African languages, and its Nano models cover 9 European and 8 African languages in files of 17 to 42 MB (video).
• VisionPsy describes images and answers questions about them, with 460 million parameters, small enough for a phone (video).
• TranscribePsy turns speech into text in 25 European languages on a laptop and 8 on a phone (video).

Each Psy model is released for research and educational purposes as open weights under Apache 2.0, which means anyone can download its trained files from Hugging Face. More Psy models are in preparation.

Fabric, the engine

QVAC Fabric is the engine that runs the models, built on llama.cpp, a popular open-source program for running AI models on ordinary computers. It also lets you fine-tune a model on the phone itself, and QVAC has shown it working on an iPhone 16 and a Samsung S25.

Genesis, open training data

QVAC Genesis is a collection of 191 billion tokens (word pieces, the unit models read) of science and maths text, generated by AI to train small models that reason well. Researchers can use it for free under a non-commercial licence (CC-BY-NC 4.0). Watch the Genesis III release video.

Who is QVAC for?

• Developers and builders put AI inside their apps with no keys to manage and no usage bill that grows with every new user.
• Anyone paying for AI gets capable models for text, images and speech at no cost, on a device they already own.
• Creators generate images and music on their own machine, free every time.
• The privacy-minded keep what they type and the files they share on their own hardware, so their privacy does not depend on a company keeping a promise.
• People off the grid still get answers where there is no signal, with no account they could be locked out of.

Projects built on QVAC, in education, health, robotics and more, are on the showcase page.

Where to start

If you want a private assistant, download QV.AC. If you want to build, run npm install @qvac/sdk and open the docs. Paolo also answered Reddit users' questions about local AI in an Ask Me Anything, and a 7-minute video covers the highlights.

FAQ