Coral Edge AI Platform

6/16/2026

Introduces the TPU Developer Hub as a new educational resource for model builders, optimizers, and developers to learn about Google Cloud TPUs. The hub provides end-to-end developer lifecycle resources covering pre-training, post-training, and inference workloads, including hardware architecture, software stack capabilities (compiler, XLA, PyTorch migration), tracing/debugging/observability (XProf tooling), parallelism/optimization strategies (Pallas kernels, KV cache offloading), and networking/security. It emphasizes open-source code recipes and deep-dive technical documentation.

6/3/2026

This post introduces the integration of Gemma 4 12B with the Google AI Edge stack, enabling on-device, agentic workflows. It highlights three key applications: Google AI Edge Gallery for local data analysis and script generation on macOS, Google AI Edge Eloquent for fully on-device voice dictation and editing with advanced instruction following, and LiteRT-LM CLI's new 'serve' command for drop-in local LLM serving compatible with standard tools and SDKs. The post emphasizes the availability of these capabilities on everyday laptops and provides benchmarks for performance.

5/19/2026

This post introduces LiteRT-LM, an optimized runtime for deploying Gemma 4 and other LLMs on edge devices. It details the underlying stack, including LiteRT, XNNPACK, and MLDriift kernels, and highlights performance gains on various hardware backends (CPU, GPU, NPU) and platforms (Android, iOS, Web). Key features include native support for Multi-Token Prediction (MTP) for up to 2.2x speedup, advanced session management for continuity and efficiency, and optimized memory utilization. The post also covers agentic workflow orchestration with Thinking Mode and constrained decoding, and announces new cross-platform interfaces for Swift (iOS) and JavaScript (WebGPU).

5/19/2026

This post announces the Beta launch of the Google Tensor ML SDK, integrating it with LiteRT for a unified developer workflow. It details the compilation process using LiteRT Torch for PyTorch or TFLite models, deployment via Play Feature Delivery and AI Packs, and running inference with LiteRT Runtime. It also highlights the Model Garden with over 100 classic ML and Generative AI models optimized for Tensor's TPU, providing examples of what developers can build, such as small language models, intelligent content creation, vision and understanding applications, and audio/accessibility tools. It also mentions specific Pixel devices supported and directs developers to documentation, sample apps, and community resources.

5/14/2026

This post details the integration of Arm Scalable Matrix Extension 2 (SME2) with Google AI Edge (LiteRT, XNNPACK, KleidiAI) to accelerate on-device AI inference on Arm CPUs. It showcases a workflow for converting PyTorch models to .tflite using LiteRT-Torch, optimizing them with Model Explorer and AI Edge Quantizer (demonstrating INT8 dynamic quantization for DiT submodule), and deploying them for high-performance inference via LiteRT's XNNPACK delegate. The post highlights over 2x speed improvements and 4x memory reduction for Stable Audio Open, while maintaining audio quality.

4/23/2026

This post details the integration of LiteRT, a cross-platform production-ready framework for on-device AI, with Neural Processing Units (NPUs). It highlights how LiteRT abstracts NPU SDK complexities, enabling CPU, GPU, and NPU acceleration across mobile, desktop, and IoT platforms. Real-world examples from Google Meet, Epic Games (Live Link Face), and Argmax Inc. (Argmax Pro SDK) demonstrate significant performance gains (e.g., 25x larger models, 2x speedup) and power efficiency achieved through NPU acceleration. The Google AI Edge Gallery App now includes NPU support for Gemma models, and the Google AI Edge Portal offers benchmarking services for mobile devices. LiteRT's cross-platform support is extended to industrial edge platforms like Qualcomm Dragonwing™ IQ8 Series and Arduino VENTUNO Q, and to AI PCs via OpenVINO™ integration with Intel® Core™ Ultra processors.

4/2/2026

Introduced Gemma 4, a family of state-of-the-art open models for on-device AI development, enabling agentic capabilities like multi-step planning, autonomous action, and offline code generation. Launched AICore Developer Preview for Android and Google AI Edge for cross-device development. Google AI Edge Gallery now features 'Agent Skills' for on-device autonomous workflows, including knowledge augmentation, content production, and model integration. LiteRT-LM enhances on-device AI with minimal memory footprint, constrained decoding, and dynamic context handling for large context windows, demonstrating high performance on edge devices like Raspberry Pi 5 and Qualcomm Dragonwing IQ8. Gemma 4 is now supported across mobile, desktop, web, and IoT platforms. Released a Python package and CLI tool for easier experimentation and pipeline development.

3/6/2026

The LiteRT stack has fully graduated into production, becoming the universal on-device inference framework. It offers 1.4x faster GPU performance than TFLite, introduces NPU acceleration, and supports a unified workflow for GPU and NPU acceleration. LiteRT also provides first-class PyTorch/JAX support via model conversion and supports cross-platform GenAI deployment for models like Gemma. TensorFlow 2.21 includes operator enhancements for lower-precision data types (int8, int16x8, INT2, INT4) for improved performance and efficiency in operators like SQRT, comparison, slice, and fully_connected. Community efforts are now exclusively focused on security and bug fixes, dependency updates, and community contributions for various TensorFlow projects.

1/28/2026

LiteRT has fully graduated into production, offering a unified on-device AI inference framework. Key advancements include 1.4x faster GPU performance than TFLite, new NPU acceleration, simplified GPU/NPU workflows across platforms (Android, iOS, macOS, Windows, Linux, Web) via ML Drift (OpenCL, OpenGL, Metal, WebGPU), asynchronous execution, and zero-copy buffer interoperability for reduced latency. NPU integration is streamlined with AOT/JIT compilation and partnerships with MediaTek and Qualcomm, achieving up to 100x faster speeds than CPU. LiteRT provides superior cross-platform GenAI support with LiteRT Torch Generative API, LiteRT-LM, and LiteRT Converter & Runtime, outperforming Llama.cpp on CPU/GPU and offering 3x NPU acceleration for Gemma models. It also supports a broad range of open-weight models and offers broad ML framework support.

12/8/2025

Introduced the LiteRT NeuroPilot Accelerator, a successor to TFLite NeuroPilot delegate, for deploying AI on MediaTek NPUs. Key features include a unified deployment workflow with AOT and on-device compilation, enhanced generative AI support (Gemma family, Qwen3), and a simplified C++ API with Native Hardware Buffer Interoperability. Detailed a 3-step deployment process and highlighted performance benchmarks for Gemma and Qwen models on MediaTek Dimensity 9500. Provided integration paths for Gemma models using LiteRT-LM and LiteRT.

11/24/2025

Introduced the LiteRT Qualcomm AI Engine Direct (QNN) Accelerator, replacing the TFLite QNN delegate. This provides a unified, simplified mobile deployment workflow for Android developers by abstracting vendor-specific SDKs and SoC fragmentation. It enables seamless deployment across supported devices with AOT or on-device compilation. The accelerator supports extensive LiteRT ops for full model delegation to the NPU, achieving state-of-the-art performance for LLMs and GenAI models. Detailed benchmarks show significant speedups over CPU and GPU, and a streamlined 3-step getting started guide is provided for AOT compilation and deployment via Google Play AI Packs.

10/15/2025

Introduces Coral NPU, a full-stack, open-source platform for edge AI. Details its AI-first hardware architecture based on RISC-V, including scalar, vector, and matrix execution units. Highlights the unified developer experience with support for IREE, TFLM, TensorFlow, JAX, and PyTorch. Mentions optimization for encoder-based architectures and small transformer models like Gemma. Outlines target applications such as ambient sensing, audio processing, and image processing. Emphasizes hardware-enforced privacy with CHERI. Announces a partnership with Synaptics for their Astra™ SL2610 processors featuring the Torq™ NPU subsystem, the first production implementation of the Coral NPU architecture.

9/24/2025

Introduces LiteRT-LM, a C++-based LLM inference framework for edge devices, as an open-source project. Details its integration into Chrome, Chromebook Plus, and Pixel Watch for on-device GenAI. Explains the LiteRT-LM architecture (Engine/Session) and optimizations (context switching, session cloning, CoW KV-cache) for multi-feature LLM deployment. Demonstrates a lightweight pipeline assembly for resource-constrained devices like the Pixel Watch. Provides access to the C++ preview and sample code.

9/9/2025

Introduced audio modality to the Google AI Edge stack with Gemma 3n, accessible via the MediaPipe LLM Inference API for Android and Web. This enables on-device speech-to-text and speech-to-translated-text capabilities for audio clips up to 30 seconds. The Google AI Edge Gallery app now includes an 'Audio Scribe' feature to demonstrate these new audio capabilities. The Gallery app has also been released on the Google Play Store for easier access.

6/26/2025

Introduced Gemma 3n, a mobile-first multimodal AI architecture optimized for on-device deployment. Key innovations include the MatFormer architecture for elastic inference and custom model sizing (Mix-n-Match), Per-Layer Embeddings (PLE) for memory efficiency by offloading embeddings to CPU, KV Cache Sharing for faster long-context processing, an advanced audio encoder based on Universal Speech Model (USM) enabling on-device ASR and AST, and the MobileNet-V5-300M vision encoder for high-performance multimodal tasks on edge devices. Gemma 3n models are available in E2B and E4B sizes, offering performance comparable to smaller traditional models with significantly reduced memory footprints.

6/24/2025

This post details how Gemini 2.5 models, including Gemini Robotics-ER, can be used for robotics and embodied intelligence. It provides practical examples and prompts for semantic scene understanding (object identification, restocking needs, gauge reading), tracking multiple objects, detecting open-ended concepts like 'spill', and generating robot arm trajectories. It also demonstrates using Gemini 2.5 with a robot control API for code generation to perform pick-and-place tasks with different strategies, and shows how in-context examples can be used for more dexterous robot control tasks like packing boxes and folding a dress.

5/20/2025

Introduced LiteRT, a new platform for on-device ML inference on mobile GPUs and NPUs. Key features include MLDrift for improved GPU acceleration with optimized data organization and workgroup optimization, NPU support co-developed with MediaTek and Qualcomm, simplified APIs for specifying target hardware accelerators (GPU, NPU), seamless buffer interoperability using TensorBuffer for zero-copy data transfers between hardware memory types, and asynchronous execution leveraging OS-level mechanisms for parallel processing across CPU, GPU, and NPUs.

5/20/2025

Introduced support for over a dozen on-device small language models (SLMs) including Gemma 3 and Gemma 3n, with Gemma 3n being the first multimodal on-device SLM supporting text, image, video, and audio inputs. Added new Retrieval Augmented Generation (RAG) and Function Calling libraries for on-device AI features. Enhanced model support through the LiteRT Hugging Face community. Introduced new quantization schemes for int4 post-training quantization, reducing model size and latency. Detailed performance metrics for Gemma 3 1B and Gemma 3n. Provided libraries for RAG and Function Calling on Android, with more platforms to follow. Introduced a Python tool simulation library for custom function calling.

3/12/2025

Introduces Gemma 3 1B, a new model size in the Gemma family, optimized for on-device deployment on mobile and web. Details the use of Google AI Edge's LLM inference API for efficient on-device processing, achieving up to 2585 tok/sec on prefill. Explains key performance optimizations including quantization-aware training (4-bit integer weights, dynamic int8 activation), optimized KV cache layouts, cached optimized weight layouts for faster loading, and GPU weight sharing for reduced memory footprint during prefill and decode phases. Provides guidance on deploying Gemma 3 1B on Android devices and in-browser, including performance metrics on specific hardware.

9/4/2024

TensorFlow Lite has been rebranded as LiteRT (Lite Runtime) to reflect its expanded vision of supporting multi-framework on-device AI models (PyTorch, JAX, Keras, TensorFlow). The core runtime functionality and .tflite file format remain unchanged. Developers using packages will need to update dependencies to new LiteRT packages. Building from source will continue to use the TensorFlow repo for now, with a move to a new LiteRT repo planned. TensorFlow Lite Support Library and Tasks are recommended to transition to MediaPipe Tasks.

6/26/2024

Introduced Model Explorer, a graph visualization tool for ML models, now publicly available as part of the Google AI Edge family. Model Explorer supports large model graphs (tens of thousands of nodes) with 60 FPS visualization, integrates with JAX, PyTorch, TensorFlow, and TensorFlow Lite, runs in Colab, and is extensible. It offers node metadata overlays for debugging performance and numeric accuracy, and side-by-side graph comparison for conversion error debugging. It is used by Waymo and Google Silicon for optimizing models like Gemini Nano.

5/29/2024

Introduced the AI Edge Torch Generative API, enabling developers to author high-performance LLMs in PyTorch for deployment on edge devices using the TensorFlow Lite (TFLite) runtime. This API offers custom transformer support, optimized CPU performance, compatibility with TFLite deployment flows, and support for models like TinyLlama, Phi-2, and Gemma 2B. It features multi-signature export for prefill and decode operations to optimize serving performance and includes LLM-specific performance optimizations such as high-performance SDPA and KVCache, leveraging TFLite's XNNPack delegate, and reducing runtime memory consumption.

5/14/2024

Introduced AI Edge Torch, a new tool that provides a direct conversion path from PyTorch models to the TensorFlow Lite (TFLite) runtime. This feature offers improved CPU performance, initial GPU support, and compatibility with over 70% of core_aten operators. It leverages TorchDynamo and torch.export for model conversion and supports quantization workflows. Performance benchmarks show competitive results against existing workflows like ONNX2TF and ONNX runtime. The release includes early adoption by partners like Shopify and hardware acceleration support through silicon partnerships.

3/7/2024

Introduces the experimental MediaPipe LLM Inference API, enabling Large Language Models (LLMs) to run fully on-device across Web, Android, and iOS platforms. Details the process of converting model weights to TensorFlow Lite Flatbuffers using the MediaPipe Python Package and integrating the LLM Inference SDK. Highlights supported LLMs (Gemma, Phi 2, Falcon, Stable LM) and their parameter sizes. Presents performance benchmarks for Time to First Token and Decode Speed, with detailed explanations of optimizations such as weight sharing, optimized Fully Connected Ops (int8/int4 quantization, ARM v9 I8MM instructions), balancing compute and memory for prefill/decode phases, custom GPU operators, pseudo-dynamism, and optimized activation data types.

2/14/2024

This post details the development of the 'We Spot Turtles!' app, which uses a machine learning model to identify sea turtle species based on facial scale and pigmentation. The app leverages Google Play and Flutter for global reach and development. The ML model's capability to assign unique codes for tracking individual turtles based on their unique features represents an advancement in species identification and conservation data collection.

10/9/2023

Introduced an on-device text-to-image generation solution for Android using MediaPipe, enabling developers to generate images directly on devices. This solution supports standard diffusion models, controllable generation via diffusion plugins (facial structures, edge detection, depth awareness), and customized generation using LoRA weights. It is compatible with Stable Diffusion v1.5 architecture and provides conversion scripts for models. LoRA fine-tuning is supported on Vertex AI.

8/18/2023

This post announces the initial release of the MediaPipe iOS SDK and an update to the Python SDK to support Raspberry Pi. It provides code examples for object detection on Raspberry Pi using Python, OpenCV, NumPy, and TFLite models, including setup instructions and model retrieval. It also details text classification on iOS using the MediaPipe Tasks API, with code snippets for creating a TextClassifier and performing classification. The post highlights the availability of these tools for on-device machine learning tasks.

7/10/2023

This post details the integration of MediaPipe Face Landmarker solution by KDDI to achieve real-time facial landmark detection and extract 52 blendshape scores for rendering realistic virtual human avatars. It provides a Python code example for using the MediaPipe Face Landmarker with blendshapes in LIVE_STREAM mode, storing blendshape data in Firebase Realtime Database, and transmitting it to Google Cloud's Immersive Stream for XR for real-time animation in Unreal Engine. It also outlines the Unreal Engine side implementation using the Firebase C++ SDK and the GameInstance Subsystem for receiving blendshape data and driving facial animations.

5/11/2023

Introduced MediaPipe Solutions, a new suite of on-device ML tools comprising MediaPipe Studio (no-code model viewing/testing), MediaPipe Tasks (low-code ML deployment libraries for web, mobile, IoT, desktop), and MediaPipe Model Maker (custom model retraining). Added nine new ML tasks including Face Landmarker, Image Segmenter, Interactive Segmenter, and Image Generator. Provided code examples for implementing gesture recognition in Android using MediaPipe Tasks and for retraining custom gesture models with MediaPipe Model Maker.

5/4/2023

This post introduces MediaPipe Model Maker as a tool for customizing ML models, highlighting the importance of data preparation, simplifying model design, expecting multiple training iterations, prototyping outside the app using MediaPipe Studio, and making incremental changes. It provides practical advice and links to relevant resources for image classification, gesture recognition, text classification, and object detection models.

5/5/2022

Announces ASUS as an OEM partner for Coral, enabling global scale for edge AI deployments. This expands the reach and availability of Coral-powered devices.

6/25/2021

This post announces the release of MediaPipe Solutions, a new collection of on-device ML tools. It introduces MediaPipe Studio for no-code model viewing and testing, MediaPipe Tasks for low-code ML deployment across vision, audio, and text, and MediaPipe Model Maker for custom model retraining. New tasks like Face Landmarker, Image Segmenter, Interactive Segmenter, and Image Generator are highlighted. Future plans include no-code model training in MediaPipe Studio.

5/26/2021

This post details the integration of MediaPipe's hand tracking capabilities into the Mirru App to enable advanced prosthesis control. It highlights how MediaPipe's on-device ML processing allows for real-time hand gesture recognition, which is then translated into commands for controlling a prosthesis. This showcases a practical application of MediaPipe for assistive technology development.

11/19/2020

This post announces new hardware and software updates for Coral in 2020, including the Coral Dev Board Mini and expanded support for TensorFlow Lite models. These updates aim to make edge AI more accessible to developers by providing more options for deploying ML models on edge devices.

10/15/2020

This post details how to build an app that detects crop diseases using TensorFlow Lite and the MediaPipe framework. It covers data collection, model training with TensorFlow, and deployment on Android devices. The process involves using a dataset of crop images, training a classification model, and integrating it into an Android application using the MediaPipe library for real-time inference. This demonstrates a practical application of on-device ML for agricultural purposes.

9/25/2020

Introduces the 3D Face Transform solution within MediaPipe, enabling real-time 3D facial mesh transformations. This solution utilizes MediaPipe's existing face detection and mesh tracking capabilities to generate a 3D mesh of the face, which can then be used for various applications like virtual try-on and avatar creation. The post details the underlying technology and provides a demo.

9/16/2020

This post announces the launch of the Coral Accelerator, a new hardware accelerator designed to bring advanced AI capabilities to edge devices. It highlights the performance improvements and expanded capabilities for on-device machine learning, enabling more complex AI models to run efficiently at the edge. The focus is on enabling developers to build powerful edge AI applications with enhanced performance and reduced latency.

8/31/2020

This post introduces instant motion tracking capabilities within MediaPipe, enabling real-time pose estimation and tracking for applications. It details the use of MediaPipe's Pose Tracking solution, which leverages machine learning models to detect and track human body landmarks in real-time from video streams. The post highlights the performance optimizations and ease of integration for developers looking to incorporate motion tracking into their applications.

8/27/2020

Introduced ML Kit Pose Detection, a new on-device pose estimation feature for mobile applications. This feature leverages the MediaPipe Pose solution to provide real-time pose tracking, including skeleton tracking, joint detection, and confidence scores. The library is designed for use in fitness, health, and interactive applications.

8/6/2020

This post introduces digital ink recognition capabilities within ML Kit, enabling developers to integrate on-device handwriting recognition into their Android and iOS applications. It details the underlying machine learning models and APIs provided by ML Kit for this purpose, allowing for real-time recognition of handwritten text, symbols, and drawings.

4/22/2020

Introduced KNIFT (Keypoint-based Neural Image Feature Tracker), a template-based feature matching system for MediaPipe. KNIFT leverages learned feature descriptors and a template matching approach to achieve robust visual tracking and recognition, even in challenging conditions like occlusions and significant viewpoint changes. It integrates with MediaPipe's graph framework for efficient on-device deployment.

3/24/2020

This post details how Alfred Camera integrated MediaPipe's on-device machine learning capabilities to implement smart camera features. It highlights the use of MediaPipe's Person Detection, Motion Detection, and Object Tracking solutions to enable real-time analysis directly on the device, reducing latency and improving privacy. The post discusses the technical implementation of these features within the Alfred Camera application, showcasing the practical application of MediaPipe for edge AI.

1/28/2020

This post introduces MediaPipe on the Web, enabling on-device machine learning directly in the browser using JavaScript. It highlights the ability to run complex ML models like pose estimation and face detection with low latency and high performance, leveraging WebGL and WebAssembly. The post details how to integrate MediaPipe into web applications, providing examples and code snippets for common use cases, and discusses the performance benefits and potential applications of on-device ML in web environments.

1/2/2020

Introduced new Coral hardware products for 2020: Coral Dev Board, Coral Dev Board Mini, Coral SoM, and Coral USB Accelerator. These products are designed to bring AI capabilities to edge devices, complementing the existing MediaPipe Solutions for on-device ML.

12/10/2019

This post introduces the integration of object detection and tracking capabilities into MediaPipe, a new framework for building machine learning solutions on device. It details how to use MediaPipe's Object Detection API and Object Tracking API to build real-time applications that can detect and track objects in video streams. The post provides code examples and discusses the underlying models and algorithms used for these tasks.

11/22/2019

This post announces Mendel Linux 4.0 for Coral, which includes updated support for TensorFlow Lite and a new API for accessing camera data. It also highlights the integration of MediaPipe Solutions, including MediaPipe Studio, Tasks, and Model Maker, for on-device ML development.

10/22/2019

Coral, an AI hardware platform for edge AI, has moved out of beta. This signifies a transition from experimental to a more stable and production-ready offering for developers looking to deploy machine learning models on edge devices.

8/6/2019

This post announces several updates to the Coral Edge AI platform and related tools. Key updates include post-training quantization support for TensorFlow Lite models, allowing for more efficient on-device inference. A new TensorFlow Lite delegate for Coral devices has been released, improving compatibility and performance. Additionally, new models have been released, including a new version of the Face Detector and a new Pose Detector.

5/31/2019

This post introduces new resources for the Coral platform, including project tutorials to guide developers in building edge AI applications, a downloadable compiler for deploying models to Coral devices, and the establishment of a new distributor program to increase hardware availability. These updates aim to lower the barrier to entry for edge AI development.

5/14/2019

This post announces new features for ML Kit, including APIs for image labeling, object detection, and barcode scanning. It also highlights the integration of ML Kit with Firebase ML, allowing for on-device model deployment and management. The post emphasizes the ease of use for developers to bring ML capabilities to their apps.

4/11/2019

This post announces updates to the Coral Edge AI platform, including a new compiler that improves model deployment efficiency on Coral hardware. It also highlights performance gains and new features for on-device machine learning, such as support for new model architectures and optimizations for specific hardware capabilities.

3/6/2019

This post announces the public release of Coral, a dedicated AI hardware platform, and its integration with the AI Edge stack. It highlights Coral's role in enabling on-device AI inference for a variety of applications, from image recognition to natural language processing. The post details the availability of Coral Dev Boards and USB Accelerators, and showcases example projects demonstrating the capabilities of local AI processing.

7/25/2018

Introduced new AIY Edge TPU boards (Dev Kit and Vision Kit) for on-device machine learning inference, enabling developers to build and deploy AI models for applications like object detection and image classification on edge devices.

5/9/2018

This post introduces ML Kit, a mobile SDK for on-device machine learning capabilities. ML Kit provides APIs for common ML tasks such as text recognition, image labeling, and barcode scanning, enabling developers to integrate ML features into their Android and iOS applications without requiring extensive ML expertise. It leverages existing Google technologies and aims to simplify the adoption of on-device ML.