Paolo Ardoino's appearance at Cantor Convergence offered a window into how far Tether's ambitions now extend beyond digital dollars. Fortune's account of the Miami keynote describes imagery drawn from Isaac Asimov's Foundation alongside decentralized phone and AI infrastructure in Africa, as Ardoino argued for a broader set of decentralized technologies.

The more concrete version of that vision is already taking shape in QVAC, Tether's local-AI platform. A new small agent model arrived during conference week, and a developer update announced on November 16 adds another step toward running more AI work on hardware people control.

What Tether means by local AI

QVAC's own explainer describes a platform for running models on phones and computers, with a software development kit for builders and a separate QV.AC assistant app. The app was described as an early public alpha in that explainer.

The central idea is that a device performs the model's computation rather than sending each request to a central AI service. That can change who has access to the information being processed and how usage is paid for. It also means the user's hardware has to be capable of running the chosen model.

Space and Time’s enterprise approach to private data and verifiable calculations addresses another part of the control question. Keeping computation local and proving a database calculation are different capabilities; businesses may need to evaluate both.

This is the thinking behind a phrase attributed to Ardoino in QVAC's explanation:

“Not your AI, not your intelligence”

— Paolo Ardoino, as quoted in QVAC's platform explainer

Local inference and a fully offline workflow are not the same thing. A model may reason on a phone while a task such as checking a live flight or sending an email still requires access to an external service. The useful question is which parts of the work remain local and what permissions the assistant needs to act.

That permission question is explored in Blockster’s coverage of AI-agent guardrails: access to information should not automatically grant authority to transfer money or change an account.

A small model designed to take actions

On November 11, Tether AI Research announced BotPsy, an open-weight model with 800 million parameters focused on tool use. Tether says it performed strongly against larger models in its own agent benchmarks; those results are company-reported evaluations, not a universal claim about every task.

The technical team's accompanying post says the weights are available under Apache 2.0 and describes tasks such as working with email and calendars on a phone or laptop. It is a specialized component for building an assistant, not proof that every desired personal-assistant workflow is finished.

For agents that pay for services, t54’s controlled credit for AI agents illustrates another layer of infrastructure. A locally running model still needs separate authorization and spending limits before it can safely handle money.

Today's update gives developers more to work with

QVAC's November 16 announcement introduces SDK 0.24, including additional image-generation models, streaming speech support and the ability for one runtime to serve multiple applications. It also reports faster vector search—the process used to find relevant information before a model answers—with results depending on the hardware tested.

The release notes are the practical counterpart to the broader Cantor vision: tools developers can inspect and try. The question is whether those tools make a local application easier to build and dependable enough to use.

This puts Tether alongside the competition covered in Blockster's Qwen4-Max report, but with a different emphasis. Qwen's launch stresses capability and task costs through model services. QVAC emphasizes control of the hardware and where computation happens.

Ardoino's ambition is expansive. The evidence to follow is more specific: usable releases, capable small models and applications that keep working under the constraints of the devices people actually own.

Reporting by Sasha Steenkamp

AIAI agentsPrivacy