- QVAC is Tether’s new AI assistant designed to run full inference directly on users’ devices, avoiding cloud dependence.
- The system achieved around 1,062 ms to first token and roughly 34.6 tokens per second in a public demo on a mid‑range laptop.
- QVAC relies on a modular architecture using the Model Context Protocol (MCP) to connect with tools like Asana without exposing data to centralized servers.
- Tether’s move into decentralized AI comes as USDT remains the leading stablecoin and the company broadens its bets on infrastructure, energy and telecom.
Tether has formally stepped into the artificial intelligence arena with the unveiling of QVAC, a new assistant designed to run entirely on local hardware, without routing queries through remote data centers. The project blends the company’s crypto‑native focus on sovereignty with a technical bet on decentralized, peer‑to‑peer AI infrastructure.
In a public demonstration shared on X, CEO Paolo Ardoino showed QVAC operating on what he described as a mid‑range laptop GPU, handling natural‑language instructions and interacting with productivity tools in real time. The assistant created and organized tasks in Asana using simple text prompts, with response times that aimed to showcase that local inference can feel just as responsive as cloud‑based services.
Tether positions QVAC as a fully local AI assistant
During the demo, terminal logs highlighted that the system was running with “100% local inference and reasoning”, meaning the model executed all computations on the user’s device rather than sending data to Tether’s servers or third‑party clouds. Metrics shown on screen included a time‑to‑first‑token of about 1,062.1 milliseconds and a generation speed near 34.6 tokens per second, figures presented to underline that a laptop‑class GPU can sustain practical workloads without massive infrastructure behind it.
This design choice aligns with the broader crypto ethos of user autonomy and data self‑custody. Instead of forwarding prompts, documents or account data to centralized clusters, QVAC keeps sensitive information confined to the device where the assistant runs. For users working with financial data, private correspondence or internal business workflows, the promise is that no external provider gains routine visibility into their interactions with the model.
Ardoino argued that as AI systems begin influencing critical decisions in finance, communication and everyday life, relying on opaque cloud stacks becomes increasingly risky. He described current AI infrastructure as fragile and intrusive, suggesting that QVAC is meant to be an alternative where the logic runs close to the user, with transparency that can be inspected once the project’s code is fully published.
In practical terms, the live test showed QVAC receiving a human‑language instruction to organize a project, interpreting the intent, then creating a main task and a related subtask in Asana without any visible delay beyond the roughly one‑second processing window. Throughout the interaction, the assistant stayed anchored to the user’s machine, rather than handing off processing to a remote inference service.
All of this is part of a larger strategy in which Tether is trying to bridge its stablecoin background with a new role in AI infrastructure. By moving away from exclusive dependence on centralized computing, the company is betting that there is room for local‑first tools that still integrate comfortably with the wider digital ecosystem.
How QVAC uses the Model Context Protocol to plug into external tools
Under the hood, QVAC relies on the Model Context Protocol (MCP), an open standard introduced in 2024 by Anthropic to define how AI models talk to external tools, services and data sources. Rather than hard‑coding every integration, the assistant can communicate with a variety of MCP‑compatible services using a common message format, which helps keep the core model lean and more maintainable.
The MCP setup follows a client-server structure in which the model functions as the client and each tool—such as task managers, calendars, file systems or databases—is represented by a dedicated MCP server. These servers expose well‑defined capabilities that the assistant can call, such as creating tasks, querying records or reading files, all via standardized protocol messages.
In the Asana example demonstrated by Ardoino, QVAC parsed the user’s instruction, mapped it to specific MCP calls and then used the Asana MCP server to generate a new task hierarchy. The assistant handled intent recognition locally, while the MCP server carried out the API‑level interaction. Depending on how users deploy it, the MCP server can also be run in an environment they control, reinforcing the privacy‑first aspirations of the project.
Before MCP emerged, connecting models to tools often required bespoke integrations, meaning each new application demanded one‑off engineering work and custom interfaces. By adopting an open protocol, QVAC aims to reduce that friction and allow developers to plug in new capabilities—like document repositories, messaging platforms or enterprise systems—without rewriting the assistant’s internal logic from scratch.
Tether has stated that QVAC will be released as open source software, opening the door to third‑party audits, forks and specialized derivatives tailored to particular industries. For developers, the combination of a local‑first assistant and an extensible MCP ecosystem could make it easier to build secure workflows where proprietary data never leaves devices under their direct control.
Modular architecture and a peer‑to‑peer AI network
Beyond the protocol layer, QVAC is built around a modular architecture that treats capabilities as discrete, composable components. Instead of packing every feature into a single monolithic application, the system is structured so that new “skills” can be added, replaced or removed without altering the underlying engine that performs inference and reasoning.
According to Tether’s description, this modularity should allow developers to iterate quickly on features such as integrations, domain‑specific reasoning modules or interface elements, while keeping the main model relatively stable. Over time, that could translate into a catalog of optional add‑ons that users pick and choose from, depending on whether they need personal productivity helpers, research tools or financial‑oriented agents.
Complementing this structure is a peer‑to‑peer (P2P) communication layer intended to link devices running QVAC without funneling traffic through central hubs. Instead of a traditional client-cloud model, QVAC nodes can talk directly to one another, sharing information or delegating tasks within a distributed mesh that lacks a single, privileged coordinator.
A report cited by Tether suggests that this network design could scale to trillions of agents and applications while avoiding single points of failure. The company uses the phrase “Infinite Intelligence swarm” to describe a future in which countless QVAC instances and related agents cooperate across a P2P fabric, collectively handling workloads that would normally be routed through a few large data centers.
From a resilience standpoint, the idea is that no single node or server becomes mission‑critical. If a subset of devices go offline, others can continue processing tasks and maintaining connections, somewhat analogous to how decentralized networks in the crypto space handle outages and routing changes without completely halting activity.
Privacy also plays a role in this design: P2P interactions can be arranged so that only the minimum necessary information is shared between agents, with computation staying as close as possible to the data source. That stands in contrast to many cloud‑hosted systems where the default is to centralize both data and computation in the same environment.
QVAC Workbench: consumer app and data foundations
The local assistant will be delivered through a dedicated application known as QVAC Workbench, which is described as a consumer‑facing environment for running AI workloads on personal devices. Rather than acting as a thin client for a remote model, Workbench is meant to provide everything needed for on‑device inference, orchestration and tool integration in one place.
Tether initially introduced the broader QVAC AI initiative in May and has continued to refine the underlying data and training infrastructure. By December, the company had expanded its QVAC Genesis II dataset to around 148 billion tokens, an effort aimed at strengthening the base model that powers the assistant. The dataset size is framed as an indicator of the breadth of information and patterns available during training.
The next major milestone for the project is a planned open‑source release, which will provide a clearer view of how the model is structured, which optimizations are in place for local hardware and how the security boundaries are implemented. That release is expected to be a key test of whether QVAC can attract a robust developer community willing to contribute new MCP servers, modules and user‑interface layers.
At the same time, Tether has been ramping up its activity in sectors that support the heavy lifting required by AI, including telecommunications, energy production and broader infrastructure investments. The company appears to be positioning itself not only as a stablecoin issuer but as an investor in the physical and digital underpinnings that advanced AI systems will increasingly rely on.
Whether QVAC Workbench will reach a mass audience remains to be seen, but the combination of local privacy guarantees and an extensible tool ecosystem gives it a distinct profile in a market still dominated by centralized, cloud‑hosted assistants.
USDT’s dominant role as Tether diversifies into AI and assets
While QVAC marks Tether’s clearest move into decentralized AI so far, the company’s core business—USDT, the largest dollar‑pegged stablecoin—continues to expand. Data from DefiLlama indicates that USDT’s market capitalization is hovering around USD $184-185 billion, keeping it firmly ahead of rival tokens in terms of circulating supply.
Alongside its international operations, Tether has rolled out USAT, a U.S.‑domiciled entity that launched in the prior month with a comparatively modest circulating supply of roughly USD $20 million. The mainline USDT token remains the company’s flagship product, servicing what Tether estimates to be approximately 530 million users worldwide, with about 30 million new users joining every quarter.
Ratings agency S&P Global has tracked changes in the composition of Tether’s reserves over time. Between September 2024 and November 2025, the share of assets categorized as “higher risk,” such as gold and Bitcoin, increased from around 17% to about 24%. Over the same period, the proportion of U.S. Treasury holdings declined from roughly 81% to 75%, signaling a gradual diversification away from an almost exclusively government‑bond‑based backing.
Even with that shift, Tether continues to hold a very substantial portfolio of U.S. Treasuries—more than USD $122 billion according to the company’s figures—representing over 83% of its total reserves. Those levels place Tether among the top twenty global holders of U.S. sovereign debt, in a bracket comparable to countries such as Germany and Saudi Arabia.
Bo Hines, who heads Tether’s U.S. operations, has suggested that the firm could soon rank among the ten largest buyers of U.S. government bonds. In parallel, Tether has become one of the biggest holders of physical gold worldwide and, in the previous year, emerged as the third‑largest shareholder of Adecoagro, a major agricultural company known as Argentina’s top producer of milk and rice.
These moves underline that Tether is gradually evolving into a broad‑based financial and infrastructure player, extending its reach beyond digital dollars into commodities, traditional debt markets and now AI. The company’s wager seems to be that control over both capital and computing resources will matter more as AI systems merge with global payments and data flows.
A new intersection of stablecoins, AI and infrastructure
Putting all of this together, QVAC can be seen as one element in a larger roadmap where stablecoins, sovereign debt holdings, energy projects and AI infrastructure feed into one another. A global user base transacting with USDT could, in theory, benefit from locally running assistants that help manage payments, analyze financial data or interface with on‑chain services, all without shipping raw information to remote clouds.
Tether frames QVAC as a way to reconcile AI’s growing influence with stricter privacy expectations. By keeping inference close to the user and offering a transparent, modular stack, the company hopes to attract individuals and organizations wary of handing their data to opaque, centralized models that may be trained or fine‑tuned on user inputs without clear boundaries.
On the technical side, the combination of MCP‑based integrations, peer‑to‑peer networking and the Infinite Intelligence swarm concept sketches out a future in which many small agents cooperate rather than depending on a few giant models hosted in proprietary clouds. Whether that architecture proves practical at scale will depend on how efficiently QVAC can run on everyday hardware and how robust the supporting network protocols turn out to be.
For now, the public demo offers a first glimpse of what Tether envisions: a privacy‑centric assistant with sub‑second local processing, tight integration with productivity tools and a design that favors composability over lock‑in. As the open‑source release approaches and more technical details emerge, developers, users and regulators alike will be watching to see how a major stablecoin issuer manages its entry into the fast‑moving world of decentralized AI.
Ultimately, Tether’s launch of QVAC and its investments in data, energy and financial assets highlight how AI, digital currencies and global infrastructure are beginning to converge, creating a landscape where the same entities that issue money‑like tokens may also operate the networks and tools that help people interact with that money in increasingly automated ways.
