Tether’s QVAC Brings Billion-Parameter AI Training to Smartphones

Tether just made billion-parameter AI models trainable on your phone. The stablecoin giant’s QVAC division launched what it claims is the first cross-platform LoRA fine-tuning framework for Microsoft’s BitNet architecture, enabling AI development on consumer hardware that previously required enterprise-grade NVIDIA systems.

The numbers are striking. A 125-million-parameter BitNet model can be fine-tuned in roughly 10 minutes on a Samsung S25 using a biomedical dataset of about 300 documents. Scale that up to a 1-billion-parameter model, and you’re looking at just over an hour on the same device. Tether’s team pushed an iPhone 16 to fine-tune models up to 13 billion parameters.

Memory Savings Change the Economics

The real story here is efficiency. Benchmarks show BitNet-1B uses up to 77.8% less VRAM than Gemma-3-1B and 65.6% less than Qwen3-0.6B during both inference and fine-tuning. That memory headroom means larger models can run on hardware that would’ve been dismissed as inadequate six months ago.

GPU inference on mobile devices clocked between two and eleven times faster than CPU performance, according to Tether’s testing on Adreno, Mali, and Apple Bionic GPUs. The framework also extends LoRA fine-tuning to non-NVIDIA hardware for the first time—AMD, Intel, and Apple Silicon are all supported.

Tether’s AI Pivot Gains Momentum

This release follows Tether’s broader push beyond stablecoins. QVAC Fabric, launched as open-source software, represents the company’s bet that decentralized AI infrastructure can compete with cloud giants. The timing is notable: Tether announced on March 25 that it hired a Big Four accounting firm for its first USDT audit, suggesting the $184 billion stablecoin issuer is working to shore up credibility across all business lines.

“When training large language models depends on centralized infrastructure, innovation becomes stagnant,” CEO Paolo Ardoino said. “The era of Stable Intelligence has just begun.”

The framework opens doors for federated learning—training AI across distributed devices while keeping sensitive data local. For developers priced out of cloud AI, that’s a meaningful shift. Whether Tether can build a viable AI ecosystem alongside its stablecoin dominance remains the open question.

Full technical documentation, including adapters and cross-platform binaries, is available on Hugging Face.

Image source: Shutterstock