- Tether launches QVAC SDK, an open-source cross-platform toolkit for running AI locally on any device.
- The BitNet LoRA framework enables fine-tuning billion-parameter models on consumer GPUs and smartphones.
- TurboQuant compresses AI memory up to 5x, allowing long-context tasks on laptops and edge devices.
- QVAC Genesis II expands the synthetic educational dataset to 148 billion tokens, improving reasoning in AI training.
Tether, the company behind the USDT stablecoin, has been quietly building a comprehensive AI ecosystem under the QVAC brand. Over the past year, it has released a series of open-source tools and datasets that aim to shift artificial intelligence away from centralized cloud servers and onto the devices people actually use. The core of this push is the QVAC SDK, a software development kit that lets developers build, run, and fine-tune AI models directly on phones, laptops, desktops, and servers without relying on remote infrastructure.
This strategy is not just about convenience. Tether CEO Paolo Ardoino has repeatedly argued that the future of AI cannot depend on a handful of hyperscale data centers. In a world where billions of autonomous machines and trillions of AI agents will coexist with humans, latency, privacy, and resilience become engineering necessities. The QVAC ecosystem is designed to make intelligence a local resource, not a rented service. The following sections break down the key components of this rapidly expanding platform.
QVAC SDK: A Universal Building Block for Local AI
Announced in April 2026, the QVAC SDK is a fully open-source, cross-platform toolkit that unifies multiple local inference engines under a single API. It is built on QVAC Fabric, a fork of llama.cpp, and integrates whisper.cpp for speech-to-text, Parakeet for voice, and Bergamot for on-device translation. Developers can write code once and have it run unchanged across iOS, Android, Windows, macOS, and Linux. This eliminates the need for platform-specific branches or cloud API dependencies, making it easier to deliver consistent AI experiences everywhere. The SDK also includes peer-to-peer capabilities powered by the Holepunch stack, enabling decentralized model distribution and even collaborative training without centralized servers.
BitNet LoRA: Training Billion-Parameter Models on Smartphones
In March 2026, Tether’s QVAC team achieved a major breakthrough by releasing the world’s first cross-platform LoRA fine-tuning framework for Microsoft’s BitNet models. These 1-bit LLMs dramatically reduce memory and compute requirements, allowing billion-parameter models to be fine-tuned on consumer hardware. Benchmarks showed that a 125M-parameter model could be fine-tuned in about 10 minutes on a Samsung S25, while a 1B model completed the same task in just over an hour on the same device. The framework also supports inference on mobile GPUs like Adreno, Mali, and Apple Bionic, with performance up to 11 times faster than CPU. Memory savings are equally impressive: BitNet-1B uses up to 78% less VRAM than comparable models, freeing up headroom for larger workloads on everyday devices.
TurboQuant: Giving Local AI Data Center-Sized Memory
One of the biggest hurdles for local AI is the memory required to handle long conversations, large documents, or entire codebases. The KV cache, which stores context during a session, can quickly balloon to several gigabytes. In June 2026, Tether released a production implementation of Google’s TurboQuant algorithm, which compresses the KV cache up to 5x while maintaining output quality. This means a laptop can now handle a 100-page legal document or hours of conversation without sending data to the cloud. The open-source release includes a full quantization pipeline and adapters for common frameworks, making it easy for developers to integrate into their own applications. TurboQuant is bundled with QVAC SDK 0.12.0, further strengthening the platform’s ability to run complex AI tasks locally.
QVAC Genesis II: The Largest Synthetic Educational Dataset
Beyond runtime tools, Tether is also investing in training data. In December 2025, QVAC released Genesis II, an expansion of its synthetic educational dataset that now totals 148 billion tokens across 19 domains. The dataset uses a novel approach called Option-Level Reasoning, which extracts structured reasoning from both correct and incorrect answers in multiple-choice questions. Independent evaluations show that models trained on Genesis II produce clearer, more accurate answers compared to those trained on other synthetic datasets. Released under a Creative Commons license, the dataset is freely available to researchers and developers, reinforcing Tether’s commitment to open, community-driven AI development.
Together, these releases paint a picture of a company that is betting heavily on decentralized, on-device intelligence. From the SDK that unifies local AI development to the memory compression that makes long-context tasks feasible, and from the fine-tuning framework that turns smartphones into training machines to the dataset that teaches models to reason, QVAC is building the foundational layer for what Tether calls the “Stable Intelligence Era.” The tools are open, the code is public, and the vision is clear: AI should belong to the people who use it, not just to the companies that own the biggest servers.