One Rust core, six SDKs, zero cloud: LLMs that ship inside your app

What it does

Xybrid is a Rust toolkit for running AI models natively on phones, desktops and game engines. It covers three families — LLMs, speech-to-text and text-to-speech — and exposes them through a consistent model().load().run() API across Flutter, Swift, Kotlin, React Native, Unity, Rust, Python and a CLI. Everything runs on-device: no cloud, no API keys.

The interesting bit

The experimental MMP feature lets you chain models in a few lines of YAML — e.g. Whisper for transcription, a small LLM for reasoning, Kokoro for speech output — and every stage’s output is retrievable, not just the final one. That’s a voice assistant defined as config rather than glue code.

Key highlights

• Single Rust core with first-class bindings for six-plus platforms, including Unity via UPM
• Curated model catalogue (Whisper Tiny, Kokoro 82M, LFM2.5 230M and others) that downloads and runs without keys
• Multi-model pipelines (MMP) defined in YAML, usable from every SDK
• Metal acceleration on iOS devices; CPU fallback on the iOS Simulator for reliability
• Apache 2.0, published on crates.io, pub.dev, Maven Central, npm and SPM

Caveats

• MMP is explicitly labeled experimental
• The model lineup skews tiny (39M–230M) — this is edge AI, not frontier inference
• Web SDK is a preview; Expo Go can’t load the native module

Verdict

Worth a look if you’re building a mobile app or Unity game that needs private, offline voice or text AI without shipping your own inference stack. Skip it if you need large models or server-side scale.