xybrid-ai/xybrid

Xybrid – On‑device AI for Apps & Games

What it is – Xybrid is a collection of native SDKs and a CLI that let you run large language models (LLMs), speech‑to‑text (ASR), text‑to‑speech (TTS) and vision‑language models directly on a device. It works offline, without cloud APIs, and supports iOS, Android, macOS, Linux, Windows and Unity.

Key capabilities

• Unified API across Flutter, Swift, Kotlin, React Native, Unity, Rust and Python.
• Multi‑model pipelines (MMP) let you chain ASR → LLM → TTS (or other combos) with a single YAML description.
• Tool calling – local LLMs can invoke user‑defined functions.
• Hardware acceleration – Metal/ANE on Apple devices, optional Vulkan on Linux; other platforms run on CPU.
• Model catalogue – dozens of ready‑to‑download ONNX, GGUF or SafeTensors models for speech, text, vision and code, all runnable offline.
• Hybrid fallback – optional cloud execution per run.

How you use it

1. Add the platform‑specific package (e.g., xybrid_flutter for Flutter, xybrid-kotlin for Android, xybrid crate for Rust, etc.).
2. Load a model by name, e.g. kokoro-82m for TTS, lfm2.5-230m for a small LLM.
3. Call model.run(...) with an Envelope containing text, audio, or image data.
4. Receive the result – usually a 24 kHz WAV audio file for TTS, or plain text for LLM output.

Example (Flutter)

final model = await Xybrid.model('kokoro-82m').load();
final result = await model.run(XybridEnvelope.text('Hello world'));
// result is a 24 kHz WAV audio buffer

Example (Multi‑model pipeline)

# voice-assistant.yaml
stages:
  - model: whisper-tiny-ggml   # speech → text
  - model: qwen2.5-0.5b       # LLM processing
  - model: kokoro-82m         # text → speech

Running the pipeline via CLI or any SDK produces a spoken answer from a spoken question, all on‑device.

Why it matters – Most mobile AI toolkits require a server or are limited to a single modality. Xybrid bundles several modalities, supports game engines, and keeps everything private by running locally. This makes it attractive for developers building chat assistants, voice‑controlled games, or any app that needs AI without internet connectivity.

License – Apache 2.0, allowing commercial use and modification.