Tether Releases QVAC SDK for Running AI Locally on Any Device

Tether bets on decentralized AI with QVAC SDK

Tether has unveiled QVAC SDK, an open-source software development kit designed to build and run artificial intelligence models directly on end-user devices. The toolkit targets hardware of every tier — from smartphones to servers — and eliminates the need for cloud infrastructure.

"Tether Launches QVAC SDK as the AI Universal Building Block that Runs, Trains, and Evolves Intelligence Across any Device and Platform" — Tether (@tether), original post

According to the company, applications built with QVAC SDK run on iOS, Android, Windows, macOS, and Linux from a single codebase, with no platform-specific adaptation required.

Why it matters

The vast majority of today's AI solutions rely on centralized cloud servers, creating bottlenecks around latency, infrastructure vulnerability, and data control concentration. Tether's approach shifts computation to the end user's device. Should this model gain widespread adoption, it could reshape the AI services market and give users greater autonomy over their data and processing power.

The case against centralized AI

Tether first introduced the QVAC platform for decentralized AI in May 2025. The company argued that the current centralized AI model cannot scale effectively due to network latency, infrastructure fragility, and control being concentrated in too few hands.

As an alternative, Tether proposed local deployment — computation happens on the user's own device, and data never leaves it. The developers highlight three core advantages:

• Stronger privacy guarantees;
• Lower latency;
• Resilience against connectivity failures and offline scenarios.

The team explained that QVAC-based applications continue functioning even with poor connectivity, making AI more practical in real-world conditions. If the internet goes down or a server farm fails, the end user's experience remains unaffected.

Under the hood: technical architecture

At the core of the SDK sits QVAC Fabric — a modified version of the lightweight llama.cpp engine for running local models. Fabric maintains full compatibility with the Llama ecosystem for text generation, embeddings, and multimodal workloads.

The platform also integrates standalone engines such as Parakeet for speech recognition and Bergamot for translation. These are unified under a single API, allowing developers to combine capabilities without altering application logic.

A notable feature is the built-in P2P layer based on the Holepunch stack, which enables:

• Distributing models without centralized servers;
• Running — and eventually training — AI across a distributed network.

Tether's vision: billions of agents, no central server

Tether positions QVAC SDK as the foundation for a new class of applications — from personal AI assistants to autonomous agents integrated with Web3 services.

Tether CEO Paolo Ardoino stated that the world is approaching a moment when billions of people will share the planet with billions of autonomous machines and trillions of AI agents. He argued that the current architecture, where every decision routes through a centralized server, simply cannot scale to meet that reality.

Previously, in December 2025, the QVAC team released an open dataset for AI model training, marking an early step in building out the platform's ecosystem.