tetherto/qvac

QVAC SDK is Tether's open-source framework for running AI models locally on any device. No cloud. No subscriptions. No data leaving your machine. Launched April 9, 2026 by the company behind the USDT stablecoin, QVAC takes an opinionated stance: "Intelligence should not be a service you rent." The SDK targets developers who want to build AI-powered apps that work offline, run on-device, and keep user data private. The technical foundation is QVAC Fabric, a fork of llama.cpp extended with additional engines. Whisper.cpp and Parakeet handle speech-to-text. Bergamot provides on-device translation. The SDK wraps everything in a type-safe JavaScript API — the main entry point for all QVAC capabilities. Cross-platform support is the headline feature. One codebase runs unchanged across iOS, Android, Windows, macOS, and Linux. Hardware support spans Apple Silicon, AMD, Intel, and Qualcomm chips. The SDK also includes QVAC Fabric's BitNet extensions — GPU-accelerated ternary inference via Vulkan, Metal, and CPU backends, with on-device LoRA fine-tuning for mobile GPUs (Adreno, Mali) and discrete GPUs. Tether also baked in decentralized model distribution primitives, allowing delegated inference without centralized infrastructure. This is the crypto DNA showing through — peer-to-peer model sharing fits Tether's broader decentralization thesis. With 71 GitHub stars and 28 forks (early days — the SDK launched 2 days ago), QVAC is still establishing its developer community. The Apache 2.0 license and Tether's backing give it credibility. Whether it can compete with the polished developer experience of cloud-based tools like ElevenCreative or Claude Code remains to be seen. Best for developers building privacy-first apps that need LLM inference, speech-to-text, and translation without cloud dependencies.