Skip to content
SPECIFICTION
Get listed

MediaPipe

developers.google.com

Free plan Open source WebWindowsmacOSLinuxiOSAndroid
Visit developers.google.com
Spec score 8.6 #8 of 37 in AI Image Recognition Software

What it is

MediaPipe gives developers libraries, tools, ready-to-run models and customizable APIs for adding AI and machine learning to applications. Its vision tasks include face and hand landmarks, gesture recognition, image classification and segmentation, object detection, and pose landmarks. Text tasks cover language detection, classification, embeddings, proofreading and summarization, with audio classification also available. MediaPipe Tasks supports Android, Web/JavaScript, Python and iOS; its cross-platform APIs are described as requiring as few as five lines of code for inference. The optimized on-device pipelines use CPU, GPU and TPU acceleration for real-time use cases. Developers can try web samples in MediaPipe Studio with text, webcam input or uploaded images. Model Maker supports transfer learning for compatible models, but it is deprecated and cannot change a model's original task. Solution API inputs are processed on-device and not sent to Google; usage and performance metrics are sent to Google. The framework can be built from source for Linux, macOS, Windows and Docker, with some configurations marked experimental.

Spec sheet

What it does
MediaPipe supplies customizable machine-learning libraries and tools for deploying AI features in applications across development platforms.Read 21 Oct 2026 · developers.google.com
Privacy
Solution API inputs such as images, video, and text are processed on-device and are not sent to Google, while usage and performance metrics are sent to Google.Read 4 Oct 2026 · developers.google.com
Support
Google directs technical questions to Stack Overflow and bug reports or feature requests to GitHub issues, which are limited to community-benefiting problems.Read 4 Oct 2026 · developers.google.com
License
The repository and Python distribution identify the software as Apache License 2.0.Read 21 Oct 2026 · pypi.org
Acceleration
MediaPipe Tasks describes optimized on-device ML pipelines with end-to-end acceleration on CPU, GPU, and TPU for real-time use cases.Read 4 Oct 2026 · developers.google.com
API extensibility
MediaPipe Tasks exposes cross-platform APIs, while the lower-level Framework uses packets, graphs and calculators to build custom pipelines.Read 21 Oct 2026 · github.com
Audio feature
MediaPipe provides an audio-classification task for categorizing sounds.Read 21 Oct 2026 · developers.google.com
Browser evaluation
MediaPipe Studio lets users test on-device models and pipelines in a browser, adjust settings and use camera, image or text inputs.Read 21 Oct 2026 · developers.google.com
Community support
The web setup guide directs users to Google AI discussions and Stack Overflow for questions and GitHub for bugs and feature requests.Read 21 Oct 2026 · developers.google.com
Company founding
Google's official history says Google Inc. was established in August 1998.Read 21 Oct 2026 · about.google
Company headquarters
LinkedIn lists Google's headquarters in Mountain View, California.Read 21 Oct 2026 · linkedin.com
Company size
LinkedIn lists Google as a public software-development company with 10,001+ employees.Read 21 Oct 2026 · linkedin.com
Current Python release
PyPI lists MediaPipe 1.1.1, released October 14, 2026.Read 21 Oct 2026 · pypi.org
Customization
Model Maker uses transfer learning to retrain compatible models with a developer’s data, and the guide suggests aiming for about 100 samples per class.Read 4 Oct 2026 · developers.google.com
Customization limits
Model Maker is deprecated and no longer actively maintained; retraining cannot change the task the original model was built to perform.Read 4 Oct 2026 · developers.google.com
Desktop and self-hosted use
The installation guide documents building and running the framework from source on Linux, macOS, Windows, and Docker, with some operating-system configurations marked experimental.Read 4 Oct 2026 · developers.google.com
Developer platforms
MediaPipe Tasks supports Android, Web/JavaScript, Python, and iOS; the Solutions guide’s availability table lists supported tasks by Android, Web, Python, and iOS.Read 4 Oct 2026 · developers.google.com
Documentation depth
The official site provides task guides, platform setup guides, API references, examples, framework concepts and build-from-source material.Read 21 Oct 2026 · developers.google.com
Framework
The low-level MediaPipe Framework is used to build efficient on-device ML pipelines and provides concepts including packets, graphs, and calculators.Read 4 Oct 2026 · developers.google.com
Generative AI API
The catalog includes an LLM Inference API and other generative-AI tasks, with LLM Inference currently listed for web.Read 21 Oct 2026 · developers.google.com
Hardware acceleration
The web BaseOptions delegate can select CPU or GPU execution, with CPU as the default.Read 21 Oct 2026 · developers.google.com
Independent technical assessment
GitNux describes MediaPipe as producing general pose and landmark outputs for motion-estimation workflows rather than application-specific event statistics.Read 21 Oct 2026 · gitnux.org
Independent use case
A Japanese academic prototype used MediaPipe Hands for real-time ten-finger landmark tracking, handedness, key mapping and typing-status visualization.Read 21 Oct 2026 · jstage.jst.go.jp
Legacy support
Support for listed MediaPipe Legacy Solutions ended on March 1, 2023, though their repository code and prebuilt binaries remain available as-is.Read 4 Oct 2026 · developers.google.com
Legacy support limit
Google ended support for listed legacy solutions on March 1, 2023; unsupported legacy binaries remain available only on an as-is basis.Read 21 Oct 2026 · developers.google.com
Limit: Android crash
Good First Issue indexes an open Android HolisticLandmarker crash involving protobuf-javalite.Read 21 Oct 2026 · goodfirstissue.org
Limit: contributor onboarding
Good First Issue reports no open beginner issues, no beginner labels and no merged pull requests in the preceding 30 days.Read 21 Oct 2026 · goodfirstissue.org
Limit: Linux memory leak
Good First Issue indexes an open issue reporting a Linux-only Python Pose Landmarker memory leak.Read 21 Oct 2026 · goodfirstissue.org
Limit: missing batch processing
Good First Issue indexes an open request for batch processing of Holistic Landmarker videos.Read 21 Oct 2026 · goodfirstissue.org
Limit: web segmentation stability
Good First Issue indexes an open Web ImageSegmenter issue describing aborts when required WebGL float-mask capabilities are unavailable.Read 21 Oct 2026 · goodfirstissue.org
Low-code APIs
The Tasks page says developers can run ML inference with as few as five lines of code using its cross-platform APIs.Read 4 Oct 2026 · developers.google.com
Metrics consent
The MediaPipe terms say app developers are responsible for obtaining informed consent about Google’s processing of MediaPipe metrics where applicable law requires it.Read 4 Oct 2026 · developers.google.com
Metrics privacy
Although task inputs stay on-device, the APIs send performance and usage metrics to Google and require developers to obtain legally required consent.Read 21 Oct 2026 · github.com
Model customization
MediaPipe Model Maker adapts compatible solution models with a developer's own data.Read 21 Oct 2026 · developers.google.com
Notable release
PyPI's release history records MediaPipe 1.1.0 on October 6, 2026 and 1.1.1 on October 14, 2026.Read 21 Oct 2026 · pypi.org
Python installation
Python applications install the package with `python -m pip install mediapipe` and import vision, text or audio task modules as needed.Read 21 Oct 2026 · developers.google.com
Python requirements
The Python setup supports Windows, macOS, Linux and 64-bit Raspberry OS, requires Python 3.9 or newer and PIP 20.3 or newer.Read 21 Oct 2026 · developers.google.com
Python wheels
PyPI lists wheels for Windows ARM64 and x86-64, Linux glibc 2.28+ x86-64 and ARM64, and macOS 11 ARM64.Read 21 Oct 2026 · pypi.org
Security-policy limit
The GitHub repository reports that it has no SECURITY.md policy and no published security advisories.Read 21 Oct 2026 · github.com
Self-hosting model
MediaPipe Tasks performs inference on the device, so applications can process images, video and text without sending those inputs to Google servers.Read 21 Oct 2026 · github.com
Supported solution platforms
The solutions table lists Android, web, Python and iOS support for many tasks.Read 21 Oct 2026 · developers.google.com
Text and audio
Text tasks include language detection, classification, embeddings, proofreading, and summarization, and an audio classification task is listed.Read 4 Oct 2026 · developers.google.com
Text features
Text solutions include language detection, text classification, text embedding, proofreading, summarization and universal embedding.Read 21 Oct 2026 · developers.google.com
Vision features
The solution catalog includes face detection, face landmarks, hand landmarks, gesture recognition, pose landmarks, object detection, image classification, image embedding, image segmentation and interactive segmentation.Read 21 Oct 2026 · developers.google.com
Vision tasks
Available vision tasks include face and hand landmarks, gesture recognition, image classification and segmentation, object detection, and pose landmarks.Read 4 Oct 2026 · developers.google.com
Web packages
Separate NPM packages exist for generative AI, vision, text and audio tasks.Read 21 Oct 2026 · developers.google.com
Web requirements
Web Tasks development requires Chrome or Safari plus a Node.js/NPM application, although CDN script tags are also supported.Read 21 Oct 2026 · developers.google.com

Plans and prices

MediaPipe Free On-device ML solutions and framework; no price listed Read 4 Oct 2026 · developers.google.com
MediaPipe open-source package Free Apache 2.0 licensed · on-device libraries and models · no hosted service pricing stated Read 21 Oct 2026 · github.com

Every MediaPipe plan and price →

Strong points and limits

Strong points

  • Offers ready-to-run models and customizable APIs
  • Lists vision, text and audio tasks
  • Tasks supports Android, Web/JavaScript, Python and iOS
  • On-device pipelines use CPU, GPU and TPU acceleration
  • Solution API inputs are processed on-device

Limits

  • Model Maker is deprecated and no longer maintained
  • Retraining cannot change a model's original task
  • Usage and performance metrics are sent to Google
  • Some desktop build configurations are experimental

Who it suits

It suits developers building AI or machine-learning features for mobile and web applications, especially those targeting on-device inference. Developers considering Model Maker should account for its deprecated status and task constraints.

MediaPipe in depth

Read the full write-up

Overview

MediaPipe is a developer toolkit for adding machine-learning inference to applications, with ready-to-run models, APIs and components for building on-device pipelines. It is best suited to teams that want to integrate several kinds of ML capability across mobile, browser or desktop software. Its breadth and acceleration options are compelling; teams seeking maintained, low-code model retraining should look elsewhere.

Key features

The task range is unusually broad for an on-device toolkit: vision includes face and hand landmarks, gesture recognition, image classification and segmentation, object detection, and pose landmarks. Text tasks cover language detection, classification, embeddings, proofreading and summarization, with audio classification as well. This makes MediaPipe a sensible base when a product needs multiple kinds of inference, though teams building only one narrowly defined image-recognition workflow may prefer a dedicated service.

MediaPipe Tasks supports Android, Web/JavaScript, Python and iOS. Its cross-platform APIs can run inference with as few as five lines of code, and optimized pipelines use CPU, GPU and TPU acceleration for real-time use cases. The short-code claim lowers the barrier to trying inference, but MediaPipe remains a toolkit to integrate into an application, not a hosted product that manages deployment for you.

MediaPipe Studio provides browser samples for text, webcam input and uploaded images, with settings such as confidence thresholds available to adjust. Model Maker can retrain compatible models using transfer learning and a developer’s data; the guide suggests about 100 samples per class. Its practical value is now limited: it is deprecated, no longer actively maintained, and cannot change the task the original model was built to perform.

Solution inputs such as images, video and text are processed on-device rather than sent to Google. Usage and performance metrics are sent to Google, however, and app developers must obtain informed consent for that processing when applicable law requires it. That distinction matters for teams evaluating privacy obligations, even when the inference itself stays local.

The lower-level MediaPipe Framework provides packets, graphs and calculators for constructing efficient on-device ML pipelines. Developers can build and run it from source on Linux, macOS and Windows, or use Docker; some operating-system configurations are experimental. Legacy Solutions remain in repository code and prebuilt binaries as-is, but support ended on March 1, 2023, making them a poor foundation for new work.

Pricing

MediaPipe is free and open source. Its MediaPipe plan costs nothing and includes on-device ML solutions and the framework. There is no free trial because the core offering is already free. The price is attractive for developers who want to build and host inference themselves; the trade-off is that the toolkit does not substitute for a managed hosted service, and deprecated Model Maker limits the appeal for teams expecting supported retraining tools.

Platforms

The listed platforms are Android, iOS, Linux, macOS, self-hosted, web and Windows. MediaPipe Tasks targets Android, Web/JavaScript, Python and iOS, while the framework installation guide documents source builds on Linux, macOS, Windows and Docker. This gives teams options from mobile and browser apps to self-hosted desktop or server environments, with experimental configurations requiring caution.

Who it's for

MediaPipe fits application developers who need configurable, on-device inference across vision, text or audio, and who can own integration and deployment. It is especially suitable when local processing and hardware acceleration matter. It is less suitable for buyers who want a hosted API, supported no-code retraining, or a turnkey application rather than libraries and framework components.

Pros and cons

Pros: Broad vision, text and audio tasks let a team build several inference features on one toolkit.

Pros: Android, web, Python and iOS APIs, plus CPU, GPU and TPU acceleration, support responsive on-device deployments across common developer environments.

Pros: Free, open-source access includes both ready-to-run solutions and lower-level pipeline tools.

Cons: Model Maker is deprecated and cannot change a model’s original task, limiting its value for ongoing customization.

Cons: Metrics go to Google, so developers may need to secure user consent depending on applicable law.

Cons: Legacy Solutions are unsupported, and some desktop configurations are experimental.

Alternatives

For a broader AI Image Recognition Software comparison, consider the category list. Choose Sighthound AI instead if its paid offering and hosted free plan—with 24-hour access and a two-minute cap—fit better than building an on-device integration. Face++ is an alternative with free API access, one API key and QPS limits that vary by API group; pick it when those API terms suit your application. Google Cloud Vision API is a cloud, per-image alternative, with each applied feature counted as a billable unit and a free allowance for the first 1,000 units. SentiSight.ai offers an offline model license with a custom quote based on model speed and a 30-day free trial, making it an option to evaluate for offline model use.

CompreFace may suit teams looking for a free, open-source Docker deployment with CPU and GPU support. InsightFace is worth considering for research and non-commercial projects; commercial model use requires a separate license. Roboflow has a free tier with 10 monthly credits, enough for roughly 30 model trainings or 80,000 inferences, for teams whose work fits that credit allowance. Yandex Vision OCR is an alternative for printed-text or license-plate recognition, with per-unit charges and no free plan.

Verdict

Choose MediaPipe if you are a developer building cross-platform, on-device inference and want a broad set of tasks with acceleration and no licensing price. Look elsewhere if you need a maintained retraining workflow, supported legacy components or a hosted service, and account for metrics consent in privacy-sensitive applications.

Verdict

MediaPipe covers a broad set of on-device AI tasks and developer platforms, with acceleration options and APIs intended to keep inference straightforward. Note the metrics transmission and the limits of deprecated Model Maker.

Where Specifiction places it

Alternatives to MediaPipeAll alternatives →