Tether, the issuer of the world's largest stablecoin USDT, has recently launched a new AI training framework. This framework aims to enable large language models to be fine-tuned on consumer hardware, including smartphones and non-Nvidia GPUs.
The framework supports cross-platform training and inference, compatible with various chips, including AMD, Intel, and Apple Silicon, as well as mobile GPUs from Qualcomm and Apple.
Tether claims that its engineers can fine-tune models with up to 1 billion parameters on smartphones in less than two hours, while smaller models can be fine-tuned in just a few minutes, with supported model sizes reaching up to 13 billion parameters.
The framework is based on BitNet, a 1-bit model architecture that can reduce VRAM requirements by up to 77.8% compared to similar 16-bit models, allowing larger models to run on limited hardware. Additionally, it supports LoRA fine-tuning of 1-bit models on non-Nvidia hardware, expanding support beyond the GPUs typically used for AI training.
The company states that these performance improvements also apply to inference, with mobile GPUs running BitNet models several times faster than CPUs. It also highlights application scenarios such as on-device training and federated learning, where models can be updated across distributed devices without sending data to a centralized server, potentially reducing reliance on cloud infrastructure.
As crypto companies gradually expand into computing and machine learning, Tether's initiative comes at a time of rising interest in Bitcoin mining and AI agents, showcasing the active trend of the crypto industry in AI and high-performance computing (HPC). AI agents, which can trade, interact with services, and execute tasks autonomously, are gaining increasing attention in the crypto space.