Tether Launches Bitnet AI Framework for Smartphones, Cutting Need for Nvidia GPUs

Tether is taking aim at Big Tech’s AI hardware moat with a framework that promises to shrink billion-parameter model training down to something your phone can handle.

Tether AI Framework Cuts VRAM Use by Over 70%, Expands Edge Computing

On Tuesday, Tether unveiled a cross-platform LoRA fine-tuning framework for Microsoft’s Bitnet models, introducing what it described as the first system capable of training and running 1-bit large language models across consumer devices, including smartphones and laptops.

The release is part of Tether’s QVAC Fabric stack and is designed to reduce the heavy compute and memory demands typically associated with artificial intelligence development, which has largely been confined to cloud providers and high-end Nvidia hardware.

By supporting heterogeneous hardware—including chips from Intel, AMD, and Apple, as well as mobile GPUs—the framework allows developers to fine-tune models locally without relying on centralized infrastructure.

In practice, that means AI workloads once reserved for data centers can now run on devices sitting in a backpack or a pocket, a shift that could lower costs and broaden access for developers across the United States and globally.

Tether said its engineers successfully demonstrated Bitnet fine-tuning on mobile GPUs, including Adreno, Mali and Apple Bionic chips, marking a first for the emerging 1-bit model architecture.

Performance benchmarks released by the company show a 125 million-parameter model can be fine-tuned in about 10 minutes on a Samsung S25 device, while a 1 billion-parameter model completes the same task in roughly 1 hour and 18 minutes on the same hardware.

On Apple devices, the company reported similar results, with a 1 billion-parameter model fine-tuned in approximately 1 hour and 45 minutes on an iPhone 16, and experimental runs pushing models up to 13 billion parameters on-device.

The framework also showed measurable gains in inference speed, with mobile GPUs delivering between two and 11 times the performance of CPUs, according to Tether’s internal benchmarks.

Memory efficiency is another key selling point, with Bitnet-1B using up to 77.8% less VRAM than comparable 16-bit models and more than 65% less than other widely used architectures, enabling larger models to run on limited hardware.

Tether said the system also enables LoRA fine-tuning on non-Nvidia hardware for the first time in this category, a move that could reduce reliance on specialized chips and cloud services while keeping sensitive data stored locally on user devices.

The company added that the approach could make federated learning more practical by allowing models to be trained across distributed devices without centralizing data, an area of growing interest in privacy-focused AI development.

“By enabling meaningful large-model training on consumer hardware, including smartphones, Tether’s QVAC is proving that advanced AI can be decentralized, inclusive, and empowering for everyone,” Tether CEO Paolo Ardoino said in a statement, adding that the company plans continued investment in on-device AI infrastructure.

The technical release, including benchmarks and implementation details, has been published through Hugging Face, signaling an effort to reach developers directly rather than gate the technology behind proprietary systems.

FAQ 🔎

• What is Tether’s new AI framework?
  Tether’s QVAC Fabric introduces a cross-platform system for training and running Bitnet AI models on consumer devices like phones and laptops.
• Can smartphones really train AI models?
  Yes, Tether’s benchmarks show billion-parameter models can be fine-tuned on devices like the Samsung S25 and iPhone 16 within hours.
• Why is this important for U.S. developers?
  It reduces reliance on expensive cloud infrastructure and specialized GPUs, lowering costs and increasing access to AI development.
• What makes Bitnet different from other models?
  BitNet uses a 1-bit architecture that significantly reduces memory usage and improves efficiency compared to traditional 16-bit models.