What QVAC actually is
Tether built an AI platform. The name is an acronym, the marketing is cosmic — underneath is a fairly specific piece of engineering.
QVAC stands for QuantumVerse Automatic Computer, which tells you nothing, and Tether’s own copy about the Stable Intelligence Era and intelligence as an element of the periodic table tells you less. Underneath the register is a concrete artifact. It is worth separating the two, because the artifact is more interesting than the language wrapped around it.
The thing itself
QVAC is an open-source JavaScript SDK, licensed Apache 2.0, that runs AI models on the device the code is running on. You install @qvac/sdk from npm. You call loadModel, then a verb — completion, transcribe, translate, generate, ragSearch — then unloadModel. The same API works on Node.js, on Bare, on Expo, which means the same code path serves a server, a desktop app, and a phone.
Underneath the JavaScript is a set of native C++ engines: a fork of llama.cpp called Fabric for language models, whisper.cpp and NVIDIA’s Parakeet for speech recognition, Bergamot for translation, stable-diffusion.cpp for images, ONNX Runtime for text-to-speech and OCR. None of these are novel. What QVAC does is put one consistent interface over all of them and handle the question of which native binary to dispatch on which platform.
If you have ever tried to ship an app that does two different AI things on two different operating systems, you know why this is worth something. Getting Whisper compiling on a Mac is a day. Adding translation is a different library with a different build system and a different API. Adding mobile is starting over. The SDK’s value is that this is somebody else’s problem now.
The part that is actually new
Everything above is convenience. It is real convenience, but it is not a reason to care about QVAC specifically rather than the several other projects that also wrap local inference engines.
The reason to care is peer-to-peer, and it comes from Tether’s other holdings. QVAC is built on Holepunch’s stack — Hyperswarm, Hypercore, Hyperdrive — the same infrastructure underneath Keet and the Pear runtime. Two things fall out of that.
The first is that models distribute peer-to-peer. The model registry is a Hyperdrive over Hyperswarm. You can run your own registry and host your own models without standing up a CDN. Weights come from other people’s devices rather than from a bucket somebody is paying egress on.
The second is delegated inference, and this is the one nobody else has. A phone can offload a model to a laptop on the same network with a config option:
loadModel({
modelSrc: ...,
modelType: "llm",
delegate: { topic, providerPublicKey, fallbackToLocal: true },
})No API gateway. No cloud account. No infrastructure at all. Two devices find each other over Hyperswarm and one runs the model for the other, end-to-end encrypted, with a fallback to local if the peer is unavailable.
This matters because it dissolves the central objection to on-device AI. The objection is that phones are small: a model that fits on a phone is a model that has been compromised down to fit on a phone. Delegated inference says the phone is not the boundary. Your devices are the boundary. The laptop in the other room is yours, the connection is encrypted, nothing goes through a third party, and the model running is as large as your largest machine can hold.
Fine-tuning on the device
The second genuinely unusual claim is training, not just inference.
Fabric combines BitNet — Microsoft’s aggressive quantization architecture — with LoRA, which reduces the number of parameters that have to move during fine-tuning. The result is that a billion-parameter model can be fine-tuned on a Samsung Galaxy S25 in roughly eighty minutes. On an iPhone 16 the team has pushed to models around thirteen billion parameters.
Sit with that for a second. Google’s Gemini Nano runs locally on your phone. You cannot change it. QVAC Fabric is the first framework that lets an ordinary smartphone learn from data that never leaves it. The privacy argument for local inference has always been slightly incomplete, because the model you are running was trained on somebody else’s corpus and adapts to you not at all. On-device fine-tuning closes that gap.
It also breaks a hardware monopoly. This kind of training has been effectively locked to NVIDIA’s specialized systems. Fabric is hardware-agnostic — Vulkan, CUDA, ROCm, whichever the GPU wants — which brings AMD, Intel, Adreno, Mali, and Apple silicon into a category they were excluded from.
What it does not do
QVAC has no browser target. There is no WASM build; if you need inference in a web page, this is the wrong tool. The Python client is on a roadmap rather than in a release, so a Python ML researcher gets nothing here today. Emulators do not work — mobile development requires physical devices. macOS on Intel is CPU-only.
And local inference is not free. It is slower than a cloud API for large models, it burns the user’s battery, and it moves responsibility for optimization onto the developer. The trade you are making is privacy, offline function, and no rate limits, against latency on heavy models and a support surface that now includes every device your users own.
Why Tether
The obvious question. A stablecoin issuer has no evident business building AI infrastructure, and the reflexive read is that this is a crypto company chasing an adjacent hype cycle.
The less cynical read is that the strategy is coherent. Tether owns Holepunch. Holepunch built the P2P stack. Tether has a wallet development kit designed so that software agents, not only humans, can hold and move value. QVAC is the intelligence layer for a stack that already had the networking layer and the payments layer. If you believe the future contains a very large number of autonomous agents transacting with each other, you would want all three, and you would want none of them to depend on a company that could switch them off.
Whether that future arrives is a separate question. The engineering is real regardless.
Where to start
The SDK documentation is at docs.qvac.tether.io, and the quickstart is a five-minute exercise. The HTTP server is OpenAI-compatible and runs on port 11434, which is Ollama’s default port — so pointing an existing client at a local QVAC server is a one-line change.
The apps in this directory are what people have built with it so far. There are not many yet. That is what an early ecosystem looks like.