Tether Unveils AI Framework Enabling Model Training on Smartphones

Tether has introduced a new artificial intelligence training framework designed to enable large language models (LLMs) to run and be fine-tuned on consumer hardware, including smartphones. This system, part of Tether's QVAC platform, utilizes Microsoft's BitNet architecture and LoRA fine-tuning to significantly reduce memory and compute requirements, thereby lowering the cost barrier for AI development.

BitNet-Based Training System

The framework is built on BitNet's one-bit neural network structures combined with LoRA fine-tuning, allowing for efficient model adjustments with low compute demands. Engineers reported that the system can train language models with up to one billion parameters on smartphones in under two hours, with smaller models completing training in minutes. The platform also supports models up to thirteen billion parameters on mobile devices.

A key feature of the system is its broad hardware compatibility, extending beyond Nvidia chips to include AMD processors, Intel architectures, Apple Silicon, and mobile GPUs from Qualcomm and Apple. This approach aims to democratize AI experimentation by moving beyond traditional high-performance computing clusters. The BitNet architecture reportedly reduces VRAM usage by up to 77.8% compared to standard 16-bit systems.

Decentralizing AI Compute

Tether's initiative aims to decentralize machine-learning capabilities, reducing reliance on centralized cloud providers. The framework enables LoRA fine-tuning on hardware outside the dominant Nvidia ecosystem, addressing a historical dependency on Nvidia GPUs for efficient tensor calculations. By allowing low-bit training methods on alternative processors, Tether seeks to remove this limitation.

The architecture also enhances inference speeds for mobile workloads, with BitNet models processing several times faster on mobile GPUs than on standard CPUs. This allows models to operate locally on devices. Furthermore, the system supports distributed machine-learning methods like federated learning, where models learn from local data on independent devices, enhancing privacy by keeping data local and only transferring model updates.

Crypto Industry's AI Expansion

Tether's move aligns with a broader trend of increasing investment in AI infrastructure within the digital asset sector. Crypto firms are repurposing computing capacity from blockchain operations for machine-learning workloads. Notable examples include Google's $3 billion deal with Cipher Mining for computing power and IREN's plans to raise $3.6 billion for AI infrastructure expansion.

Companies like HIVE Digital Technologies and Core Scientific have also expanded their high-performance computing services, with Core Scientific securing a $500 million loan from Morgan Stanley. Additionally, developers are integrating AI agents with blockchain, with Coinbase launching wallet tools for on-chain transactions and Alchemy enabling agents to access blockchain data. Identity networks, such as World (co-founded by Sam Altman), are exploring AI's role in digital verification through tools like AgentKit.

Tether's framework enters this converging space of computing resources, machine learning, and blockchain, allowing developers to integrate training tools into distributed applications and local devices without central server reliance. Future developments will depend on developer adoption and performance testing across diverse consumer hardware.