Tether has launched a new AI training framework under its QVAC platform, enabling large language models to be fine-tuned on consumer hardware including smartphones and non-Nvidia GPUs.
The system leverages Microsoft’s BitNet architecture and LoRA techniques to reduce memory and compute requirements, lowering the cost and barriers to AI development.
The framework supports a wide range of hardware, including AMD, Intel, Apple Silicon and mobile GPUs, allowing cross-platform training and inference.
Tether said models with up to one billion parameters can be fine-tuned on smartphones in under two hours, with smaller models processed in minutes and larger models supported on mobile devices.
The BitNet-based approach reduces VRAM requirements by up to 77.8%, enabling more efficient model training and faster inference on limited hardware.
The company highlighted use cases such as on-device training and federated learning, which reduce reliance on centralised cloud infrastructure.
The launch reflects a broader trend of crypto firms expanding into AI infrastructure, alongside investments from mining companies and partnerships involving major tech and financial firms.