Tether, the company behind the largest stablecoin by market capitalization, USDT, has unveiled a new artificial intelligence training framework designed to enable large language models to be fine-tuned using consumer-grade hardware, including smartphones and GPUs that are not produced by Nvidia.

In an announcement on Tuesday, the company said the system is part of its QVAC platform and is built on Microsoft’s BitNet architecture combined with LoRA techniques. These technologies significantly reduce memory usage and computational demands, potentially lowering both costs and hardware barriers associated with AI development. The framework is designed to operate across multiple platforms, supporting a wide range of chipsets, including those from AMD, Intel and Apple Silicon, as well as mobile GPUs developed by Qualcomm and Apple.

Tether stated that its engineers successfully fine-tuned models with up to one billion parameters on smartphones in less than two hours, while smaller models could be trained in a matter of minutes. The system also supports models with up to 13 billion parameters on mobile devices.

Efficiency gains and decentralized AI potential

Built on BitNet’s 1-bit model architecture, the framework can reduce VRAM requirements by as much as 77.8% compared to conventional 16-bit models, according to Tether. This allows more complex models to run on hardware with limited resources. The system also enables LoRA-based fine-tuning on non-Nvidia GPUs, broadening compatibility beyond the hardware typically required for AI training.

Tether highlighted that performance improvements extend to inference as well, with mobile GPUs capable of running BitNet-based models several times faster than CPUs. The company also emphasized potential applications such as on-device training and federated learning, where models can be updated across multiple devices without transferring data to centralized servers, reducing reliance on cloud infrastructure.

Crypto and AI convergence accelerates

Tether’s entry into AI infrastructure reflects a broader trend among cryptocurrency companies expanding into computing and machine learning, particularly as activity increases in both Bitcoin mining and AI agent development.

In September, Google acquired a 5.4% stake in Cipher Mining as part of a long-term $3 billion agreement focused on AI data center capacity. Later in December, Bitcoin mining firm IREN announced plans to raise approximately $3.6 billion to support AI-related infrastructure.

Momentum has continued into 2026. In February, HIVE Digital Technologies reported record revenue of $93.1 million, driven by growth in its AI and high-performance computing operations. In March, Core Scientific secured a $500 million loan facility from Morgan Stanley, with the option to expand it to $1 billion.

At the same time, AI agents—autonomous systems capable of executing transactions and interacting with digital services—are gaining traction within the crypto ecosystem. In October, Coinbase introduced wallet infrastructure that allows AI agents to perform onchain transactions. More recently, Alchemy launched a platform enabling agents to access blockchain data services using USDC on Base. Also in February, Pantera and Franklin Templeton joined Arena, a testing platform developed by Sentient for enterprise-grade AI agents.

On Tuesday, World, the identity network co-founded by OpenAI CEO Sam Altman, released AgentKit, a toolkit that allows AI agents to verify their connection to a unique human identity using World ID, while also enabling payments through the x402 micropayments protocol.

Sources:

https://tether.io/news/tethers-qvac-launches-worlds-first-cross-platform-bitnet-lora-framework-to-enable-billion-parameter-ai-training-and-inference-on-consumer-gpus-and-smartphones/

https://cointelegraph.com/news/tether-launches-ai-training-framework-for-smartphones-and-consumer-gpus