Xybrid

Xybrid is a runtime and platform for running machine learning models on a user's device, with automatic fallback to the cloud when the device cannot handle a request. It is aimed at developers building mobile, desktop, game and edge applications who want local inference without writing separate code for each platform. One API covers text generation, transcription and embeddings across iOS, macOS, watchOS, tvOS, visionOS, Android, Flutter, Unity and Linux or edge targets, with SDKs for Swift, Kotlin, Dart, C# and Rust. A router decides per request whether a call runs on device, in the cloud or in a hybrid mix, based on model size, available memory, battery and the developer's latency budget. The code sample on the home page loads a small language model from the registry in a few lines of Swift. The platform side adds over the air model updates with canary, regional or fleet wide rollout and one click rollback, evaluations across chips, OS versions and quantization levels, per model optimization, and telemetry and cost analytics. The vendor states that data stays on the device by default, telemetry is opt in, and SOC 2 is in progress. The runtime is released under the Apache 2.0 licence, with source on GitHub, a model registry, an in browser demo and a documentation site. Developers can request an API key. The site does not publish prices for the platform or any paid plan.