Free and open source
QVAC is the building block of intelligence you can snap into almost anything: a phone, a laptop, a robot, a wearable, a home appliance. Each one becomes smart on its own, with no internet and no cloud behind it. And it is completely free.
Who it's for
Not just AI labs and big tech. If you use a phone or a laptop, it is already for you.
Developers and builders
A solo builder adds transcription, RAG and text generation to their note-taking app in an afternoon. No API keys to manage, no per-user inference bill as it grows. It runs on each user’s own device, on iOS, Android and desktop.
Start buildingWhy on-device
Offline
Inference runs with no connection. The model lives on the device, not behind an API.
Private
No data leaves your hardware. No third-party sees your prompts - private by architecture.
Free
No API bills, no per-token pricing, no rate limits. Apache 2.0, open source.
Yours
Run it on a phone, a laptop, a server - or all three over peer-to-peer. You own the stack.
The ecosystem
Five products that work together - run models, train them, fine-tune them and ship them in apps, all on-device.
Run it
A single API for local, decentralized AI. One install in JavaScript or Python, 10+ AI tasks, on every device - text, vision and audio through one interface.
Explore SDKThe SDK
A single API for local, decentralized AI. One install in JavaScript or Python, 10+ AI tasks, on every device - text, vision and audio through one interface.
Chat, completion and structured output.
Vectorize text for search and RAG.
Retrieval-augmented answers over local docs.
Speech-to-text, fully on-device.
Translate between languages offline.
Generate images locally, no cloud.
Natural speech synthesis on-device.
Extract text from images and scans.
LoRA fine-tuning, even on mobile.
Understand images with vision models.
Upscale to high resolution on-device.
Millisecond labels with confidence.
And many more, with new capabilities in every SDK release
See the full task listThe data
The data to train your own models. An open dataset of synthetic STEM reasoning, built so independent researchers can train reasoning-capable models of their own.
The engine
Fine-tune on the device. The engine that adapts models to your own data, right on the device, even a phone, so the model learns where your users are, not in a data center.
Fine-tune on your device. Adapt a model to your data.
Runs on any GPU. Vulkan and Metal, desktop and mobile.
Even on a phone. A dynamic tiling algorithm fits tight mobile GPU memory.
The models
The models QVAC builds. Each one is trained for a single domain and sized to run on hardware you already own, so the intelligence behind your product does not depend on someone else's servers staying up.
Built for the device, not shrunk to fit it. The phone in your pocket is the constraint they are designed around.
Open weights, every model. Apache 2.0, published with the benchmarks behind them.
Small because they are specialized. A model that only has to be good at one thing does not need to be enormous.
How it compares
One on-device SDK against typical cloud AI and local runtimes.
YesYesYesYesYesYesYesYesYesYesYesYesIn the wild
Choose a path
Pick a path. We send you straight to the right part of the ecosystem.
Build
Install the SDK and run your first model on-device in minutes. 10+ AI tasks, one API, iOS, Android and desktop.
Connect
Start the local HTTP server and point your coding agent or harness at it. Official plugins for OpenCode and OpenClaw.
Create
Advance the art with Genesis or fine-tune on-device with Fabric and LoRA, even on a phone.
Explore
Browse open models tuned to run on your device, from small LLMs to vision and medical models like MedPsy.
Not sure which one fits?
Take the guided setupCommunity
Questions, releases and work in progress, out in the open with the people building on QVAC.
Ask questions and compare notes with the team and other developers building locally.
Get started
One install in JavaScript or Python. 10+ AI tasks. Private, offline and free - on the hardware you already own.
npm install @qvac/sdk