AI training is moving from cloud servers to everyday hardware as on device ai reaches flagship smartphones and consumer GPUs.
Tether unveils QVAC Fabric for local AI training
Tether, issuer of the USDT stablecoin, has introduced QVAC Fabric, a new AI training framework designed to run large language models on smartphones and consumer GPUs using Microsoft’s BitNet architecture and LoRA optimization techniques.
The company says QVAC Fabric can cut memory usage by up to 90% versus standard 16-bit models. Moreover, this reduction allows models that would normally require data centers to run directly on phones, laptops, and non-Nvidia GPUs.
Tether reports that its engineers fine-tuned models with up to 1 billion parameters on smartphones in under two hours, while smaller models required just minutes. That said, the framework is not limited to small networks and can scale significantly.
Running billion-parameter models on iPhone and Android
On flagship devices such as the iPhone 16, Pixel 9, and Galaxy S25, the team pushed fine-tuning to models as large as 3.8 billion parameters. On Apple’s latest phone specifically, they report reaching 13 billion parameters.
The framework supports a broad range of hardware, including AMD, Intel, and Apple Silicon chips, as well as mobile GPUs from Qualcomm and Apple. However, it is explicitly designed to operate without relying on Nvidia’s ecosystem, highlighting a push toward more accessible AI infrastructure.
According to Tether, mobile GPUs running BitNet-based models can operate between 2 and 11 times faster than CPU-only configurations. This performance gap underlines why mobile-focused architectures are becoming critical for local model training.
Federated learning and privacy-focused AI
One of the main use cases highlighted by Tether is federated learning, an approach where AI models are updated across many devices without sending personal data to centralized servers. In practice, this lets users personalize models locally while keeping sensitive information stored on their own hardware.
Moreover, this method reduces dependence on large cloud providers and could lower costs for smaller labs and independent developers. Tether has open-sourced the QVAC platform’s code on GitHub, inviting the community to experiment with and extend the framework.
Tether positions QVAC Fabric as a way to make on-device ai more practical at scale, especially for applications that demand strict data privacy. However, its success will depend on how quickly developers adopt the tools in real-world products.
Crypto firms race to build AI infrastructure
Tether’s launch fits into a wider shift across the crypto sector, where companies rooted in digital assets are investing heavily in AI and high-performance computing. In September 2024, Google acquired a 5.4% stake in Cipher Mining as part of a $3 billion agreement linked to AI data center capacity.
Bitcoin miner IREN announced plans in December 2024 to raise around $3.6 billion for AI infrastructure expansion. Moreover, in February 2025, HIVE Digital Technologies reported record revenue of $93.1 million, driven by AI and high-performance computing growth.
In March, Core Scientific secured a $500 million loan facility from Morgan Stanley, with an option to expand it to $1 billion. That said, these investments show how miners and infrastructure providers are diversifying beyond pure bitcoin operations.
Web3 meets AI agents and identity tools
On the same day Tether revealed QVAC Fabric, World, the identity project co-founded by Sam Altman of OpenAI, launched AgentKit. The toolkit enables AI agents to verify real human links using World ID and to initiate payments via a micropayments protocol.
Also in February, Alchemy introduced a system that lets AI agents access blockchain data services using USDC on the Base network. Moreover, this integration signals a growing convergence between smart agents, identity layers, and on-chain settlement.
Overall, QVAC Fabric underscores how Tether and other crypto-native companies are positioning themselves at the intersection of digital assets, AI research, and decentralized infrastructure, potentially reshaping how advanced models are trained and deployed at the edge.
Source: https://en.cryptonomist.ch/2026/03/18/on-device-ai-qvac-fabric/
