Tether launches the world's first mobile "billion-level AI training framework," enabling LoRA training on iPhones and Samsung devices.

In the field of artificial intelligence (AI), training powerful models has always been synonymous with "burning money," heavily reliant on expensive NVIDIA systems or cloud computing power. However, stablecoin giant Tether is attempting to rewrite this rule with technology. On March 17, Tether Data, Tether's technology division, announced the official launch of the world's first cross-platform BitNet LoRA tuning framework for its QVAC (QuantumVerse Automatic Computer) platform.

The core value of this technology lies in its ability to enable AI models with "billion-parameter scale" to learn in a personalized way directly on the mobile phone in everyone's pocket.

The magic of 1-bit architecture: enabling phones to achieve "big results with small resources"

This groundbreaking advancement is built upon Microsoft's BitNet 1-bit LLM architecture. Through optimizations using QVAC Fabric, the BitNet model's memory footprint and computational load have been reduced to extremely low levels. According to the announcement, the framework not only supports common NVIDIA GPUs but also achieves full compatibility with Intel, AMD, Apple M-series chips, and mobile devices such as Adreno (Android), Mali, and Apple Bionic GPUs.

This means that AI that could previously only run in data centers can now be fine-tuned using "Low-Rank Adaptation (LoRA)" on your mobile phone. Tether points out that this technology allows edge devices to process models that are "twice as large" as traditional Q4 quantized models, demonstrating the ultimate memory advantage.

Real-world test data revealed: The astonishing speed of the Samsung S25 and iPhone 16

In their announcement, the Tether engineering team shared exciting test data demonstrating the framework's practical capabilities on modern mobile phones:

• 125 million parameter model: Fine-tuning a dataset of 300 biomedical documents on a Samsung S25 takes only about 10 minutes .
• 1 billion (1B) parameter model: The same fine-tuning task took 1 hour and 18 minutes on the Samsung S25 and 1 hour and 45 minutes on the iPhone 16 .
• Extreme Challenge: The development team successfully ran a model with up to 13 billion (13B) parameters on the iPhone 16 for fine-tuning, pushing the physical limits of mobile devices.

Say goodbye to API keys and create 100% private personal AI.

Tether CEO Paolo Ardoino has consistently emphasized: "If you need an API key to use AI, then it doesn't really belong to you." QVAC's core philosophy is "Local-first."

Through the BitNet LoRA framework, users can enable AI to learn directly from local emails, notes, and messages without uploading any data to cloud servers. This not only eliminates enterprises' concerns about the misuse of sensitive data but also breaks the current situation where AI development is limited to the monopoly of a few giants. Currently, QVAC Fabric LLM has been released as open-source software (Apache 2.0 license) and provides pre-selected adapters on Hugging Face, allowing developers worldwide to immediately launch this revolution in edge computing.

Related reports

Rakuten unveils "Japan's largest AI model" with great fanfare; its social media discovery engine is DeepSeek V3.

Musk: If I win the lawsuit against OpenAI, I'll donate it all to charity! The court has ruled that the maximum damages could reach $109 billion.

143 million players scanned street views to play Pokémon GO, inadvertently helping Niantic build the largest real-world visual dataset for AI.