Tether Unveils AI New Framework: Mobiles Can Train Billion-Parameter AI Models

On March 17, stablecoin issuer Tether announced that its AI platform QVAC Fabric has launched the world’s first cross-platform LoRA fine-tuning framework for Microsoft’s BitNet (1-bit large language model). This framework enables billion-parameter-scale LLMs to be trained and inferred on everyday hardware—including laptops, consumer-grade GPUs, and smartphones. Officially, the framework drastically cuts the VRAM and computing power requirements for AI model training, supporting Intel, AMD, Apple Silicon, and a range of mobile GPUs (such as Adreno, Mali, and Apple Bionic). In testing: - A roughly 1.25-billion-parameter BitNet model fine-tunes in ~10 minutes on a Samsung S25. - A 10-billion-parameter model takes ~1 hour 18 minutes on the S25, and ~1 hour 45 minutes on an iPhone 16. - The team even successfully fine-tuned a 13-billion-parameter model on an iPhone 16. For performance: - BitNet models see 2–11x faster inference speeds on mobile GPUs than on CPUs. - BitNet-1B’s VRAM usage during inference and fine-tuning is reduced by up to 77.8% vs. a 16-bit model. Paolo Ardoino noted the technology is built to reduce reliance on large-scale cloud computing and specialized AI hardware, enabling local-device AI training. It also lays groundwork for emerging models like decentralized AI and federated learning.