This repository contains official and community contributed sample applications, model recipes, agent skills and utilities for LiteRT (formerly known as TensorFlow Lite), Google's open source, high-performance on-device machine learning framework and LiteRT-LM, a specialized orchestration layer for running LLMs with LiteRT, unlocking maximum performance and efficiency.

Note Please access the interactive web page with a collections of demos there at: https://google-ai-edge.github.io/litert-samples/

The samples demonstrate different API paradigms (LiteRT CompiledModel API and legacy Interpreter API, Tensor API, LiteRT-LM) and provide end-to-end model conversion and deployment pipelines.

• 🐱 Streaming TTS (KittenTTS nano): Added a tiny (15M-param, 32 MB) streaming text-to-speech Android sample — dynamic-length LiteRT graphs, sentence-level streaming playback, live TTFA/RTF metrics (samples/litert/text_to_speech_streaming/).
• 🎙️ Speech Recognition (ASR): Added end-to-end Automatic Speech Recognition sample using the CompiledModel API.
• 📸 PhotoTalk Sample App: Added multimodal sample app combining LiteRT vision processing with LiteRT-LM audio/text generation (samples/litert/phototalk_sample_app/).
• 🗣️ Qwen3-TTS & Qwen3 ASR: Added model recipes, conversion scripts, and Tensor API implementations for Qwen3-TTS and Qwen3 ASR.
• 🎨 Bonsai Image 4B: Added text-to-image diffusion model sample with Python inference and conversion tools (models/bonsai/bonsai_image_4b/).
• 🤖 Agent Skills & Utilities: Added four lifecycle agent skills (skills/: conversion, quantization, on-device verification, app scaffolding), a GPU conversion toolkit (utilities/litert_gpu_toolkit/), and shared Kotlin helpers (utilities/common/).

All runnable sample applications and interactive playgrounds are organized under samples/:

• samples/litert/: Standard samples using the LiteRT CompiledModel API. Designed for modern hardware acceleration (GPU/NPU) and asynchronous execution.
  • Samples: Speech Recognition, PhotoTalk, Text-to-Speech, Image Generation (text-to-image), Image Segmentation, Image Classification, Digit Classification, Qualcomm NPU acceleration (Gemma, MobileNet, Fast VLM), Google TPU sample app.
• samples/litert_model_zoo/: One Android app that runs 29 vision and audio models through the LiteRT CompiledModel API, each downloaded on demand from Hugging Face.
• samples/litert_interpreter/: Legacy samples using the Interpreter API.
  • Samples: Broad compatibility examples for Android, iOS, and Python (Image Classification, Object Detection, Image Segmentation, Audio Classification).
• samples/litert_lm/: High-level Engine samples for Large Language Models (LLM/SLM).
• samples/end_to_end/: Complete full-system pipelines (e.g. ImageNet model conversion, preprocessing, and classification).
• samples/tensor_api_playground/: Interactive Web/WASM playground demonstrating LiteRT Tensor API capabilities directly in the browser (Gemma 3, Image Segmentation, Mandelbrot, Game of Life).

Contains standalone model conversion scripts, export recipes, and model-specific utilities. Many are working in process. Start with the model conversion cookbook: the step-by-step instructions for taking a Hugging Face checkpoint to a verified .litertlm bundle or .tflite graph, with the recipes as worked examples.

• utilities/common/: Shared Kotlin helpers for Android samples (camera pipeline, audio capture, CompiledModel runner, image/tensor and math helpers).
• utilities/litert_gpu_toolkit/: Pre-conversion patches that rewrite common PyTorch patterns into forms the LiteRT GPU delegate accepts, plus a post-conversion checker.

Custom AI agent skills that carry a model through the LiteRT deployment lifecycle, in order — see skills/README.md for the full index:

• litert-conversion-workflow/: Hugging Face LLM / VLM checkpoint → verified .litertlm bundle for LiteRT-LM.
• gpu-clean-conversion/: PyTorch / Hugging Face model → GPU-resident LiteRT model.
• accuracy-safe-quantization/: Quantize (fp16 / int8 / int4) without losing accuracy.
• on-device-verification/: Prove the converted model on the actual device.
• compiled-model-app-scaffolding/: Build an Android app around the verified model.
• litert-compiled-model-migration/: Android TFLite → LiteRT CompiledModel V2 with NPU JIT acceleration and zero-copy buffers.

• Android: Android Studio (latest stable version).
• iOS: Xcode (latest version).
• Python: Python 3.9+ and pip install ai-edge-litert.
• Web / WASM: Modern browser with WebGPU / WebAssembly support.

1. Navigate to samples/litert/<sample_name>.
2. Ensure you have a device with a supported NPU/GPU (e.g., modern Pixel, Samsung, or Qualcomm/MediaTek devices).
3. Follow the specific setup instructions in the sample's README.md.

Visit "Interactive Web" tab, or open samples/tensor_api_playground/index.html or run index.html at the repository root via a local HTTP server.

• LiteRT Overview: ai.google.dev/edge/litert
• LiteRT-LM Overview: ai.google.dev/edge/litert-lm
• CompiledModel API Guide: LiteRT for Android
• Model Conversion: Convert models to LiteRT
• LiteRT CLI: CLI for end-to-end journey

Contributions are welcome!

1. Read CONTRIBUTING.md.
2. Fork the repo and create a branch.
3. Submit a Pull Request.

Apache License 2.0. See LICENSE for details.

Disclaimer: This is a sample repository maintained by Google. It is provided "as is" without warranty of any kind.