Tether has unveiled an AI training framework that lets developers fine‑tune large language models on consumer hardware, including smartphones and non‑Nvidia GPUs. The system, integrated into Tether’s QVAC platform, combines Microsoft’s BitNet architecture with LoRA techniques to slash memory and compute requirements and lower the hardware barrier for model development.
The cross‑platform framework supports training and inference across AMD, Intel and Apple Silicon CPUs, plus mobile GPUs from Qualcomm and Apple. Tether says its engineers fine‑tuned models of up to 1 billion parameters on smartphones in under two hours, with smaller models taking only minutes. The company also reports work extending support to models as large as 13 billion parameters on mobile devices.
Built around BitNet’s 1‑bit model design, the framework can reduce VRAM needs by as much as 77.8% compared with equivalent 16‑bit models, enabling larger models to run on limited hardware. It also enables LoRA fine‑tuning for 1‑bit models on non‑Nvidia hardware, moving beyond the traditional reliance on Nvidia GPUs for training workloads.
Tether highlights inference gains as well: mobile GPUs running BitNet models reportedly deliver multiple‑times faster performance than CPUs. That performance, combined with reduced memory use, opens practical scenarios such as on‑device training and federated learning, where models are updated across distributed devices without centralizing user data. These approaches can reduce latency, cloud costs and dependence on centralized infrastructure.
Tether’s expansion into AI tools comes as some crypto and blockchain companies diversify into compute and machine learning. Recent industry moves include Google taking a 5.4% stake in Cipher Mining under an AI data center deal, Bitcoin miner IREN planning multibillion‑dollar capital raises for AI infrastructure, HIVE Digital Technologies citing record revenue driven by AI and HPC, and Core Scientific arranging a large credit facility for data center growth.
The sector is also seeing growth in AI agents—autonomous programs that interact with services and execute transactions. Examples include Coinbase introducing wallet infrastructure for AI agents to make on‑chain transactions; Alchemy launching a system for agents to access blockchain data services using USDC on Base; Pantera and Franklin Templeton joining Arena to test enterprise AI agents; and World (co‑founded by Sam Altman) releasing AgentKit so agents can verify links to unique humans via World ID and make payments with the x402 micropayments protocol.
Tether’s announcement underscores an industry shift toward more distributed, device‑centric AI workflows and broader hardware diversity for model training. Readers are encouraged to verify technical claims and performance figures independently.