xybrid-ai/xybrid

Xybrid – On‑device AI for apps & games

What it is – Xybrid is an open‑source SDK that lets you run large language models (LLMs), speech‑to‑text (ASR) and text‑to‑speech (TTS) locally on a wide range of platforms. It ships a single, consistent API that works in Flutter, Swift, Kotlin, Unity, Rust, Python and the web, and it can load models in ONNX, GGUF or SafeTensors formats.

Why it matters – All inference happens on the device, so no cloud service or API key is required after the initial model download. This gives you privacy, offline capability, and the ability to swap models at runtime without republishing the app.

Core capabilities

Capability	Details
Speech‑to‑Text	Whisper‑tiny, Wav2Vec2‑base and other ONNX/ggml models.
Text‑to‑Speech	Kokoro‑82M, KittenTTS‑Nano, NeuTTS‑Nano – output is a 24 kHz WAV file.
LLM inference	Hundreds of models (LFM2.5, SmolLM2, Gemma, Llama 3.2, Qwen 3.5, Bonsai 27B, etc.) with tool‑calling and chain‑of‑thought support.
Vision‑language	Compact VLMs (LFM2‑VL, LFM2.5‑VL).
Multi‑model pipelines (MMP)	Chain ASR → LLM → TTS (or any sequence) via a tiny YAML description.
Hardware acceleration	Metal + Apple Neural Engine on iOS/macOS, optional Vulkan on Linux, CPU on other platforms.
Hybrid/cloud fallback	Optionally route a run to a remote service if the device can’t handle it.
Telemetry (opt‑in)	Usage stats can be sent behind an API key.

Platforms & language bindings

Platform	Binding	Package manager
Flutter	`xybrid_flutter`	pub.dev
iOS / macOS (Swift)	`xybrid` (Swift Package)	Swift Package Manager
Android (Kotlin)	`xybrid-kotlin`	Maven Central
Unity	`ai.xybrid.sdk`	OpenUPM / Git URL
Rust	`xybrid` crate	crates.io
Python	`xybrid` binding (source)	pip (via source)
Web (preview)	JavaScript SDK	npm (preview)
CLI	`xybrid` executable	script installer

Getting started (example in Flutter)

# pubspec.yaml
dependencies:
  xybrid_flutter: ^0.6.0

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

The same pattern works in Kotlin, Swift, C# (Unity) and Rust – you load a model by its ID, feed an Envelope (text or audio), and receive the processed audio.

Model catalogue

Xybrid ships a built‑in registry (see xybrid.ai/models) that includes:

ASR – Whisper‑tiny (39 M), Wav2Vec2‑base (95 M)
TTS – Kokoro‑82M (82 M, 24 voices), KittenTTS‑Nano (15 M), NeuTTS‑Nano (120 M)
LLMs – from 230 M up to 27 B parameters, covering multilingual, reasoning, tool‑calling and multimodal variants.
Vision‑language – compact VLMs (450 M‑3 B). You can also add your own model (experimental) by providing a model_metadata.json that describes the file format and preprocessing steps.

How it differs from similar projects

Feature	Xybrid	Ollama	llama.cpp	ONNX Runtime
Mobile (iOS/Android)	✅	❌	❌	✅
Unity integration	✅	❌	❌	❌
Multi‑model pipelines	✅	❌	❌	❌
All three modalities (ASR + LLM + TTS) in one SDK	✅	❌	❌	❌
Runs in‑process, no server	✅	❌	✅	✅
Offline‑first, no cloud required	✅	✅	✅	✅

License & community

License – Apache 2.0 (permissive, commercial‑friendly).
Docs – https://docs.xybrid.dev
Discord – https://discord.gg/YhFHHkhbad
GitHub – Issues, pull requests, good‑first‑issues, and a CONTRIBUTING guide are provided.

TL;DR

Xybrid gives developers a unified, offline‑first way to embed speech recognition, text‑to‑speech and large language model inference into mobile apps, desktop software and Unity games, with ready‑made bindings for the major languages and hardware‑accelerated runtimes.