LiteRT: High-Performance On-Device Machine Learning Framework | Google AI Edge

LiteRT is Google's on-device framework for high-performance AI deployment.

Efficient conversion, runtime, and optimization for on-device AI.

Get Started with LiteRT Deploy LLMs with LiteRT-LM Build from LiteRT GitHub Download LiteRT models from HF

Build once, run everywhere

Cross-platform deployment

Deploy across mobile, web, desktop, and IoT.

Generative AI

Engineered with superior support for LLMs, Diffusion, and more generative models.

Hardware acceleration

CPU, GPU, and NPU hardware acceleration with SOTA performance across leading edge platforms.

Multi-framework conversion

Direct export and quantization from PyTorch, TensorFlow, and JAX into .tflite format.

Built on the battle-tested foundation of TensorFlow Lite

Trusted by prominent Google apps

250K+ applications, billions of global users

Migrate your TensorFlow Lite app

What you can build with LiteRT

Local GenAI & assistants

Run open-weight LLMs like Gemma on-device with LiteRT-LM, without a server round trip.

View Gemma 4 sample app →

Real-time image segmentation

Run real-time, multi-class image segmentation from the camera on CPU, GPU, or NPU.

View segmentation demo →

Live speech recognition

Transcribe speech on-device with open-weight ASR models like Whisper, Moonshine, and Parakeet on CPU, GPU, or NPU.

View speech sample app →

On-device vision-language

Run the FastVLM vision-language model on-device with NPU acceleration on Qualcomm devices.

View FastVLM demo →

Streaming text-to-speech

Synthesize natural speech fully offline, with playback starting after the first sentence while the rest streams in.

View TTS sample app →

IoT & edge robotics

Deploy lightweight models, such as anomaly detection and sensor fusion, to microcontrollers and embedded devices.

Explore microcontroller docs →

Deploy with LiteRT

Streamline your AI workflow from training to on-device deployment.

1. Convert

2. Quantize

3. Deploy & Accelerate

Convert

Convert PyTorch, JAX, or TensorFlow models to .tflite, or start from pre-trained .tflite models.

Explore conversion

Quantize

Use the LiteRT quantizer to quantize your models post-training.

Explore optimization

Run

Deploy your model with LiteRT and pick the optimal accelerator for your app.

Run your first inference

Choose LiteRT Runtime or LiteRT-LM

Use LiteRT Runtime to run models with CPU, GPU, and NPU acceleration. For LLMs, use LiteRT-LM, which builds on LiteRT to manage conversations, the KV cache, and streaming output.

High-performance runtime for ML workloads. Tailored for hardware-accelerated execution across CPU, GPU, and NPU backends with advanced features such as zero-copy buffer interoperability and async execution.

Explore LiteRT Supported Silicon →

LiteRT-LM for LLM

LiteRT-LM is an orchestration layer built on the LiteRT Runtime that powers the LLM execution loop—seamlessly integrating hardware-accelerated inference, multimodal inputs, constrained decoding, and agentic tool calling.

Explore LiteRT-LM Gemma 4 Guide →

Install LiteRT Packages

Python

# Platform: Linux, macOS, Windows (Python 3.9+)
pip install ai-edge-litert

Android (Kotlin)

// Platform: Android (API 24+) | In app/build.gradle.kts
// GPU and NPU acceleration are built in; select via Accelerator in Options.
implementation("com.google.ai.edge.litert:litert:2.2.0")

iOS (Swift)

// Platform: iOS 15+, macOS 12+
// Swift Package Manager support is in progress.
// For a production iOS build today, see the LiteRT-LM tab.

Web (JS)

# Platform: Web (Browsers & Node.js)
npm install @litertjs/core

C++

# Platform: Linux, Android NDK, macOS, Windows
# In CMakeLists.txt (link LiteRT C++ SDK)
target_link_libraries(${PROJECT_NAME} PRIVATE litert_cc_api)

CLI Tools

# Platform: Linux, macOS, Windows | Model conversion & benchmarking
pip install litert-cli-nightly

Python

# Platform: Linux, macOS, Windows | LiteRT-LM Python SDK
pip install litert-lm-api

Android (Kotlin)

// Platform: Android (API 24+) | In app/build.gradle.kts
implementation("com.google.ai.edge.litertlm:litertlm-android:latest.release")

iOS (Swift)

// Platform: iOS 15+, macOS 12+ (Apple Silicon GPU)
// In Package.swift dependencies:
.package(url: "https://github.com/google-ai-edge/LiteRT-LM", from: "0.16.0")

// In target dependencies:
.product(name: "LiteRTLM", package: "LiteRT-LM")

Web (JS)

# Platform: Web (WebGPU required)
npm install @litert-lm/core

C++

# Platform: Linux, Android NDK, macOS, Windows
# Option A: prebuilt C API (headers + shared libs), no build required.
# Download litert_lm_c_api-*.zip from github.com/google-ai-edge/LiteRT-LM/releases
target_include_directories(${PROJECT_NAME} PRIVATE ${LITERT_LM_SDK}/include)
target_link_libraries(${PROJECT_NAME} PRIVATE ${LITERT_LM_SDK}/lib/linux_x86_64/liblitert-lm.so)

# Option B: build LiteRT-LM from source, then link the C++ aggregate target.
target_link_libraries(${PROJECT_NAME} PRIVATE LiteRTLM::Aggregate)

CLI Tools

# Platform: Linux, macOS, Windows | On-Device LLM runner & test harness
pip install litert-lm

Full LiteRT-LM API guide → Quick start, streaming, GPU, multimodal input, and tool use in Python, Kotlin, Swift, and JavaScript.

Supported Platforms: Android • iOS • Web • Linux • macOS • Windows

LiteRT Guide → CLI Guide → LiteRT-LM Guide →

Samples, models, and demos

Explore open-source code samples, pre-trained models, and interactive demos.