Tether has released QVAC, an open-source software development kit that runs and fine-tunes AI models entirely on local hardware, as part of the stablecoin issuer's increasingly expensive push to become something more than a dollar-peg factory.
Tether has released QVAC, an open-source software development kit for running and fine-tuning AI models entirely on local hardware — no cloud, no remote inference, no central model registry. The announcement on 9 April is the clearest signal yet that Tether's AI division is not a side project. It is the company's second front, and the one Paolo Ardoino keeps pointing at when he talks about what Tether will look like in five years.
QVAC stands for QuantumVerse Automatic Computer, a name that does the SDK no favours. The substance is more interesting. The kit wraps a modified fork of llama.cpp, the C++ runtime that has become the de facto standard for running open-weight language models on consumer hardware, under Tether's own component called QVAC Fabric. On top of that sit best-in-class local engines that Tether has either forked or bundled: whisper.cpp and Parakeet for speech-to-text, and Bergamot for on-device translation. It supports text generation, embeddings, image and audio workloads, and inference of models that ship as standard GGUF files.
What separates QVAC from the growing family of llama.cpp wrappers is its distribution layer. The SDK includes primitives for what Tether calls delegated inference: if a given device cannot run a model, it can borrow compute from a peer on the same network, with the work split across multiple local nodes rather than routed to a central GPU farm. If one of those peers drops out, the others pick up the load. It's a kind of mesh approach to inference that is closer to BitTorrent than to OpenAI — and it's deliberately designed to keep data local even when compute can't be.
The business case for Tether doing this is less obvious than the technical case. Tether is the largest stablecoin issuer in the world, with 185 billion dollars in USDT circulating and roughly 6.2 billion dollars in net profit reported for 2024. It does not need to win an AI land war to make money. What it does need is something to do with the enormous cash pile that now sits in its reserve account — and infrastructure plays that reinforce the brand without requiring new regulatory permissions are exactly the kind of capital deployment that Ardoino has said he prefers. An open-source SDK is cheap to maintain and expensive for rivals to match; the marginal cost of every additional developer who uses it is effectively zero.
The strategic angle is that QVAC makes sense for a company with one unusual liability: trust. Tether has spent a decade answering questions about whether its reserves are real, and only recently engaged KPMG for its first full audit. Any product that ships without a third-party backend is, by construction, one that doesn't rely on Tether for anything other than the code. That is a useful asset to stockpile when the core business is a stablecoin.
Technically, the SDK lands in a crowded field. Ollama, LM Studio, llamafile and a handful of hardware-specific runtimes already do most of what QVAC's language-model components offer on their own. Where Tether has a real chance to differentiate is the fabric layer — the ability to split a model across devices on a local network. That is something none of the incumbent tools handle cleanly, and it matters because the on-device AI ceiling is bounded by RAM and memory bandwidth, not by compute. A framework that lets a laptop and a phone share a 30-billion-parameter model without sending tokens to a remote server is a genuine capability.
What QVAC does not solve is the model problem. The SDK runs whatever you give it, which means the quality of the end-user experience still depends on whoever is shipping the weights — Meta, Mistral, DeepSeek, or whichever lab is currently winning the open-weights race. Tether is not training its own foundation model, at least not publicly, and the SDK is agnostic about which weights you load into it. That keeps the project honest but caps the product's upside; QVAC will rise and fall with the open-weights community around it, not independently of it.
For a company that announced a 500 billion dollar private raise less than 48 hours before releasing the SDK, the timing is telling. Tether is not asking developers to buy anything. It is asking them to build on top of a framework that happens to carry its name — which, if the open-source community rewards it, is a cheaper form of brand-building than any amount of paid marketing.
