Major Breakthrough in AI Training Thresholds
In the AI model development space, training large language models has typically required expensive hardware and cloud resources, concentrating the technology within a handful of major institutions.
(Source: Tether)
Tether's recent launch of QVAC Fabric introduces a new LoRA fine-tuning framework designed specifically for BitNet (a 1-bit large language model). This breakthrough significantly lowers computing and memory requirements, enabling ordinary users to participate in AI model training.
A core advantage of QVAC Fabric is its broad hardware compatibility. The framework runs on a wide range of devices, including:
-
Laptops
-
Consumer-grade GPUs (Intel, AMD, Apple Silicon)
-
Smartphones (including various mobile GPUs)
This means AI models are no longer confined to data centers or specialized hardware—they can now be trained and run directly on everyday devices.
Training Large Models on Mobile Phones
One of the standout features of this technology is its ability to fine-tune models on mobile devices.
For example:
-
On a Samsung S25 (Adreno GPU), a 125M-parameter model can be fine-tuned in about 10 minutes
-
On the same device, a 1B-parameter model takes about 1 hour and 18 minutes
-
On an iPhone 16, a 1B-parameter model requires roughly 1 hour and 45 minutes
The team has even succeeded in running models with up to 13B parameters on a smartphone, highlighting the rapidly growing AI capabilities of mobile hardware.
Compared to conventional models, the BitNet architecture demonstrates clear advantages in performance and resource efficiency:
Memory Usage
-
Reduces VRAM usage by up to approximately 77.8% compared to 16-bit models
-
Provides greater operational capacity, supporting larger models and personalized applications
These improvements make it much easier to deploy AI applications on edge devices.
Moving Beyond NVIDIA and Cloud Dependence
Traditional AI training has been heavily reliant on NVIDIA hardware and cloud services. QVAC Fabric breaks this dependency by enabling 1-bit LLM LoRA fine-tuning on non-NVIDIA hardware—including AMD, Intel, Apple Silicon, and mobile GPUs such as Adreno and Mali. This shift not only lowers costs but also fosters a more decentralized AI development landscape.
Advancing Decentralization and Privacy Protection
Another key benefit of QVAC Fabric is its support for data privacy and distributed learning:
-
Model training can be performed locally, eliminating the need to upload sensitive data
-
Facilitates federated learning
-
Reduces dependence on centralized infrastructure
These features point toward a safer and more scalable path for the future AI ecosystem.
Tether’s Vision for the Future of AI
Paolo Ardoino notes that AI will play a pivotal role in future society, and its advancement should not be monopolized by a small group of resource holders. He emphasizes that excessive reliance on centralized architectures for AI training not only stifles innovation but also threatens the overall stability of the ecosystem. Enabling AI to operate on personal devices is therefore a crucial step toward broader adoption.
Summary
Tether’s QVAC Fabric represents not just a technological innovation but a potential transformation in AI development models. By lowering hardware barriers and strengthening cross-platform capabilities, large language models are steadily moving out of data centers and into everyday devices. As these technologies continue to evolve, AI is poised to shift from centralized resources to a more open, decentralized, and widely accessible future.
