Tether Now Lets You Fine-Tune AI Models on Your iPhone

TLDR

• Tether has introduced a new AI training framework, QVAC Fabric, designed to operate on smartphones and consumer-grade GPUs.
• By leveraging Microsoft’s BitNet architecture and LoRA methods, the system can decrease memory consumption by as much as 90%.
• Fine-tuning of models containing up to 13 billion parameters is achievable on an iPhone 16.
• The framework is compatible with chips from AMD, Intel, Apple Silicon, and mobile processors from Qualcomm and Apple.
• This initiative marks a step in Tether’s wider transition from a stablecoin provider to a company focused on AI and infrastructure.

---

(SeaPRwire) –   Tether, the firm responsible for the USDT stablecoin, has launched a novel AI training framework that enables the fine-tuning of large language models on common devices such as smartphones and GPUs not made by Nvidia.

JUST IN: Tether’s QVAC unveils BitNet LoRA framework, bringing large-scale AI training to consumer GPUs and smartphones. pic.twitter.com/kiQOZp744L

— SwanDesk (@SwanDesk) March 17, 2026

This framework is a component of Tether’s QVAC platform and incorporates Microsoft’s BitNet architecture combined with a method known as LoRA. This combination significantly reduces the memory and computational demands for training AI models, enabling them to function on hardware that typically lacks the capability.

According to Tether, the system can lower memory needs by up to 90% relative to conventional 16-bit models. This efficiency allows more substantial AI models to operate on mobile phones, laptops, and affordable graphics cards.

The company reports that its engineers successfully fine-tuned models with as many as 1 billion parameters on smartphones in less than two hours. Models of a smaller scale were completed in mere minutes.

Using high-end devices such as the iPhone 16, Pixel 9, and Galaxy S25, the team managed to fine-tune models with up to 3.8 billion parameters. Specifically on the iPhone 16, they accomplished this with models reaching 13 billion parameters.

The framework is designed for broad hardware compatibility, supporting AMD, Intel, and Apple Silicon processors, in addition to mobile GPUs produced by Qualcomm and Apple.

Tether states that mobile GPUs utilizing BitNet models can perform between 2 and 11 times quicker than systems relying solely on a CPU.

On-Device AI Without the Cloud

Tether emphasizes federated learning as a key application. This technique enables AI models to be updated across numerous devices without transferring personal information to central servers, thereby diminishing the need for cloud-based infrastructure.

This method permits users to customize AI models directly on their own devices, ensuring their data remains local instead of being sent to an external service.

The source code for QVAC is now publicly available on GitHub, providing an opportunity for developers and smaller research teams to utilize and expand upon it.

Crypto Companies Are Investing in AI Infrastructure

Tether’s announcement aligns with a larger pattern within the cryptocurrency sector, where numerous firms originally focused on digital assets are now investing heavily in AI and computational infrastructure.

As part of a $3 billion agreement related to AI data center capacity, Google acquired a 5.4% share in Cipher Mining in September 2024.

In December 2024, the Bitcoin mining company IREN revealed its intention to secure approximately $3.6 billion to fund AI infrastructure projects.

HIVE Digital Technologies announced a record $93.1 million in revenue for February 2025, attributing the growth to its expansion in AI and high-performance computing.

Core Scientific obtained a $500 million loan facility from Morgan Stanley in March, which includes an option to increase the total to $1 billion.

Coinciding with Tether’s news, World, the identity network co-founded by OpenAI’s Sam Altman, released AgentKit. This toolkit helps AI agents confirm a connection to a real person via World ID and conduct transactions using the x402 micropayments protocol.

In a separate February development, Alchemy introduced a system that enables AI agents to utilize USDC on the Base network to access blockchain data services.