Tether introduces AI framework enabling smartphone-based language model training
The stablecoin giant Tether has revealed an innovative AI framework allowing large language models to operate and undergo fine-tuning on consumer-grade hardware and smartphones, diminishing dependence on cloud-based systems.
The company behind USDT, the world's most dominant stablecoin measured by market capitalization, has rolled out an innovative AI training framework designed to enable fine-tuning of large language models on everyday consumer hardware, ranging from smartphones to GPUs manufactured by companies other than Nvidia.
As revealed in an announcement made on Tuesday, this system forms part of Tether's QVAC platform and leverages Microsoft's BitNet architecture alongside LoRA techniques to minimize both memory and computational demands, which could substantially decrease the financial costs and hardware prerequisites associated with AI model development.
The newly introduced framework provides cross-platform capabilities for both training and inference operations across various processor types, encompassing AMD, Intel and Apple Silicon chips, in addition to mobile GPU solutions from Qualcomm and Apple.
The company reported that its engineering team successfully fine-tuned models containing up to 1 billion parameters on smartphone devices in less than two hours, while smaller-scale models required only minutes, with the framework's capabilities extending to accommodate models as expansive as 13 billion parameters on mobile platforms.
Developed using BitNet, which employs a 1-bit model architecture, the framework can slash VRAM requirements by as much as 77.8% when compared with equivalent 16-bit models, as stated by the company, thereby enabling more substantial models to function on hardware with limited resources. Additionally, it facilitates LoRA fine-tuning on hardware beyond Nvidia's GPUs for 1-bit models, broadening compatibility beyond the graphics processing units conventionally utilized for artificial intelligence training.
According to the company's statements, the performance improvements also apply to inference operations, with mobile GPUs executing BitNet models at speeds multiple times faster than CPUs. The company further highlighted potential applications including on-device training and federated learning, where model updates can occur across distributed devices without requiring data transmission to centralized server systems, which could potentially decrease dependency on cloud-based infrastructure.
Crypto companies expand into AI, from mining infrastructure to autonomous agents
Tether's strategic entry into AI infrastructure development arrives amid a broader movement of cryptocurrency companies venturing into computational resources and machine learning domains, with this activity intensifying across the Bitcoin mining sector and through the emergence of AI agents.
During September, Google acquired a 5.4% ownership stake in Cipher Mining through a $3 billion, 10-year agreement connected to AI data center infrastructure capacity. The following December saw Bitcoin mining operation IREN announce plans to secure approximately $3.6 billion in funding dedicated to AI infrastructure development.
This sector transformation has persisted into 2026. During February, HIVE Digital Technologies disclosed record-breaking revenue totaling $93.1 million, driven by expansion in its AI and high-performance computing (HPC) business segments, while Core Scientific obtained a $500 million loan facility from Morgan Stanley during March, featuring an option to increase it to $1 billion.
The cryptocurrency mining industry's strategic shift toward AI and HPC infrastructure occurs alongside the growing prominence of AI agents—autonomous software programs capable of conducting transactions, interfacing with various services and performing tasks—which are building momentum throughout the cryptocurrency ecosystem.
In October, Coinbase unveiled wallet infrastructure designed to enable AI agents to perform onchain transactions. The following month, Alchemy introduced a system granting agents access to blockchain data services through USDC payments on Base. February also witnessed Pantera and Franklin Templeton joining Arena, a platform developed by Sentient for evaluating enterprise-focused AI agents.
On Tuesday, World, the identity verification network co-created by Sam Altman of OpenAI, released AgentKit, a development toolkit enabling AI agents to confirm their association with a unique human individual through World ID verification capabilities while conducting financial transactions via the x402 micropayments protocol.