QVAC – Modular AI Agents for Privacy, Performance and Control (qvac.tether.dev)

Tether today unveiled QVAC, a modular, privacy-first platform for running AI agents locally and peer-to-peer — an explicit alternative to cloud-hosted models and centralized gatekeepers. QVAC provides a cross‑platform SDK (Linux, macOS, Windows, Android, iOS) that lets developers load, stream-complete from, and unload models with a single codepath. The bundled example shows a JavaScript API (import, loadModel, completion, unloadModel) pulling quantized Llama 3.2 1B instruct models (Q4_0 gguf) via HTTP or pear:// URLs and serving token streams in real time, enabling offline, low-latency inference and on-device learning while keeping data encrypted and never sent to the cloud. For the AI/ML community this shifts more control to endpoints: it lowers operational dependence on cloud providers, mitigates data-exfiltration risks, and enables applications that require offline operation or strict privacy guarantees. Technical implications include embracing quantized, small-footprint models for edge efficiency, native P2P distribution of model artifacts, and support for multiple use cases from the same interface. Trade-offs remain—device compute, storage, update/patch mechanisms, and content moderation/safety when models run outside centralized control—which will shape adoption and governance choices as local AI gains traction.