- Tether introduces QVAC SDK, an open-source, cross‑platform toolkit to run AI directly on local devices.
- The SDK builds on QVAC Fabric, a fork of llama.cpp, and unifies text, speech, vision and translation in a single API.
- Peer‑to‑peer features powered by Holepunch enable decentralized model distribution and delegated inference.
- CEO Paolo Ardoino calls centralized AI “a dead end” and positions QVAC as a foundation for a new "stable intelligence" era.
Tether, the company best known for issuing the USDT stablecoin, has unveiled QVAC SDK, an open-source software development kit aimed at taking artificial intelligence out of the cloud and bringing it back onto users’ own devices. The launch signals a strategic move that goes far beyond digital currencies and places the firm at the crossroads of blockchain, peer‑to‑peer infrastructure and next‑generation AI tooling.
At the core of this announcement is a clear stance from Tether CEO Paolo Ardoino, who has described centralized AI as “a dead end”. In his view, routing every AI decision through distant data centers creates limits in terms of latency, resilience and control. QVAC SDK is presented as Tether’s attempt to offer a different path, where intelligence runs locally, is shared across devices in peer‑to‑peer fashion and is less dependent on a handful of big technology providers.
Tether’s bet on local-first AI and user control
According to Tether, QVAC SDK is designed as a local‑first platform that executes and fine‑tunes AI models directly on smartphones, laptops, edge servers and industrial hardware, instead of relying on centralized cloud infrastructure. The same codebase is meant to behave identically across iOS, Android, Windows, macOS and Linux, reducing the usual friction of targeting multiple operating systems.
This design means that everyday capabilities such as writing assistance, translation, voice transcription, smart search or personal finance helpers can run directly on the device, with data processed on‑board rather than continuously sent back and forth to remote servers. Tether emphasizes that this approach not only improves privacy, but also allows applications to stay usable when connectivity is weak, unreliable or temporarily unavailable.
The company frames QVAC as part of a broader shift in how AI should be delivered. Instead of seeing intelligence solely as a cloud service that users effectively rent, Tether argues that AI should be something users actually own and host on their own hardware. This mirrors the philosophy behind cryptocurrencies and peer‑to‑peer networks, where intermediaries are minimized and participants retain greater control over their assets and data.
From a practical standpoint, the local‑first angle is coupled with an explicit focus on resilience. Tether suggests that applications built on QVAC SDK are meant to keep functioning even if a data center goes offline, a region loses connectivity, or a cloud provider faces outages. In that sense, the toolkit is pitched as an alternative to architectures where a single point of failure can bring entire services down.
A single codebase across iOS, Android, Windows, macOS and Linux
One of the most emphasized promises of QVAC SDK is its cross‑platform behavior: developers write once and deploy across multiple environments without maintaining separate branches for each operating system. Tether states that applications built with the SDK can run the same logic on mobile phones, desktops and servers, all using a unified interface.
This approach aims to simplify a common pain point in AI development today, where teams often juggle different toolkits, bindings and platform‑specific workarounds. By exposing capabilities through a single abstraction layer, QVAC SDK is intended to allow faster iteration and a more consistent user experience, regardless of the device in use.
For end users, this could translate into AI‑powered tools that feel familiar and behave consistently whether accessed from a smartphone, a laptop or an industrial edge node. Tether mentions scenarios like personal productivity, day‑to‑day financial planning, on‑device document analysis and assistive technologies that continue to work reliably even in low‑bandwidth environments.
Behind this proposition is the idea that reducing fragmentation makes it easier for developers to target “billions of humans, billions of autonomous machines and a trillion AI agents”, a phrase Tether uses to describe the scale of the environment it expects to see in the coming years. In such a landscape, a unified development layer is portrayed as a practical necessity rather than a convenience.
QVAC Fabric and the technical stack behind the SDK
Under the hood, QVAC SDK is built on QVAC Fabric, a fork of llama.cpp that provides compatibility with the llama.cpp model ecosystem. This foundation enables support for a range of tasks, including text generation, embeddings and multimodal workloads, all tailored for local execution on consumer and professional hardware.
In addition to llama.cpp‑based models, the SDK integrates several established local inference engines such as whisper.cpp, Parakeet and Bergamot. Whisper.cpp and Parakeet are used for speech‑to‑text workloads, while Bergamot targets on‑device machine translation. Each of these components is surfaced through a common API, allowing developers to mix and match capabilities without reshaping their application logic.
Through this single interface, QVAC SDK exposes a wide range of functions, including text completion, embedding generation, computer vision, optical character recognition (OCR), text‑to‑speech, speech‑to‑text and translation. Tether indicates that the set of features is expected to grow as more modules and models are integrated into the ecosystem over time.
The unification of these tools is intended to address the complexity of today’s AI landscape, where teams frequently depend on separate libraries, cloud APIs and platform‑specific pipelines for each type of task. By contrast, QVAC’s architecture aims to provide a consistent development surface for all supported modalities, making it easier to swap models, add new capabilities or optimize performance without rewriting business logic.
Peer-to-peer distribution and delegated inference with Holepunch
Beyond local execution, a defining element of QVAC SDK is its peer‑to‑peer layer powered by the Holepunch stack. Instead of pulling large AI models exclusively from centralized servers, the toolkit includes primitives for distributing models and delegating inference tasks across a decentralized network of devices.
This P2P functionality allows participants to share model artifacts, push updates and even hand off specific inference workloads to peers, all without relying on a single hosting provider. Tether describes this as a way to avoid bottlenecks and single points of failure, while also aligning with the broader peer‑to‑peer ethos present in many open internet projects.
Looking ahead, the company plans to extend these capabilities to support P2P swarms for training, fine‑tuning and decentralized inference. The goal is for these behaviors to remain transparent to developers and to operate identically on every supported platform, keeping the complexity of orchestration hidden behind the SDK’s interface.
In practice, such a model could enable scenarios where devices collectively host and improve AI systems, rather than depending on centralized training clusters alone. This would fit with Tether’s positioning of QVAC as an infrastructure layer for distributed intelligence, where networks of agents can evolve over time through local and shared computation.
Paolo Ardoino’s critique of centralized AI
Alongside the technical details, Tether is using QVAC SDK to articulate a broader critique of today’s cloud‑heavy AI ecosystem. Paolo Ardoino has argued that centralized AI cannot scale to a world of billions of autonomous machines and trillions of software agents, due to inherent constraints on latency, resilience and control.
He points specifically to the speed of light as a hard physical limit on how quickly remote servers can respond to a massive, globally distributed set of AI requests. From this perspective, always sending decisions back to a central data center introduces unavoidable delays, which become more significant as more devices request real‑time or near real‑time responses.
Ardoino also emphasizes the risks of single points of failure and concentration of power. If critical AI systems depend on a handful of providers, outages or policy changes at those companies can have far‑reaching consequences. In that light, local and decentralized approaches are framed as more robust, because intelligence is spread across many nodes rather than sitting behind a few API endpoints.
These comments resonate with a growing debate in the technology sector about who should control data, models and computational resources. While large cloud infrastructures offer scale and convenience, they also raise concerns related to surveillance, lock‑in and governance. QVAC SDK, as described by Tether, is intended to sit on the other side of that spectrum, betting on open‑source tools, on‑device processing and peer‑to‑peer interactions.
From stablecoins to the “era of stable intelligence”
Historically, Tether has been predominantly associated with USDT, the world’s largest stablecoin by market capitalization. With QVAC SDK, the company is openly signaling that its ambitions extend into AI infrastructure and research, an area it groups under the Tether Data and QVAC initiatives.
The firm describes QVAC as an advanced AI research program focused on open, decentralized and adaptive systems capable of living and learning on any device. Within this narrative, QVAC SDK is framed as a “universal building block” for what Tether calls the “Era of Stable Intelligence” – a future in which humans, machines and software agents operate side by side at massive scale.
To support this vision, Tether has pledged substantial investments into the QVAC open‑source ecosystem over the coming months and years. Planned developments include specialized toolkits aimed at robotics and brain‑computer interfaces, along with additional components needed to build complex distributed applications.
Tether Data, the unit behind QVAC, presents itself as part of a broader push toward peer‑to‑peer connectivity, privacy and transparency. Its stated objective is to enable communication and data exchange without unnecessary intermediaries, using decentralized infrastructure that is designed to be efficient and resilient. QVAC fits into that picture as the AI layer that runs on top of these networks.
Public communications from Tether and posts shared on social platforms describe QVAC SDK as a foundational toolkit that can execute, train and evolve AI across devices and platforms. Further documentation and technical resources are being made available through the official portal at qvac.tether.io, where developers can explore code, examples and integration guides.
Across all these elements, QVAC SDK is positioned less as a single product and more as a long‑term infrastructure bet: a way to build and distribute intelligence that lives on the edge, travels over P2P networks and prioritizes local ownership of data and models, rather than relying exclusively on centralized clouds and proprietary APIs.
