MediaPipe
Google's open-source framework for cross-platform, on-device ML pipelines: vision, text, audio, and LLM inference.
About
MediaPipe is Google's open-source framework for building machine learning pipelines that run on-device. It offers cross-platform APIs for Android, iOS, web, desktop, and edge devices, with prebuilt Tasks covering computer vision (face, hand, pose, object detection, segmentation, gesture recognition), text, audio classification, and on-device LLM inference, all backed by swappable pre-trained model bundles and hardware acceleration.
The project has two layers: MediaPipe Tasks, high-level libraries for ready-to-use ML tasks with minimal code, and the MediaPipe Framework, a low-level graph-based pipeline builder (packets, graphs, calculators) with a C++ core for custom workloads. MediaPipe Model Maker fine-tunes models on custom data, MediaPipe Studio provides no-code benchmarking and prototyping, and the stack underpins Google's AI Edge Gallery and LiteRT on-device AI efforts. MediaPipe grew out of Google's internal ML infrastructure, with roots going back to 2012.
What it can do
- Cross-platform on-device ML tasks: vision, text, audio, LLM inference
- Hardware-accelerated real-time pipelines on CPU, GPU, and TPU
- Swappable pre-trained and custom .tflite/.task models
- Model Maker for fine-tuning models on custom data
- MediaPipe Studio for no-code benchmarking and prototyping
Setup
pip install mediapipe for Python, or add the Android, iOS, or web SDK; start from the MediaPipe Tasks quickstart on the Google developers site.
Prerequisites
Python 3.9+ (or an Android/iOS/web toolchain); a .task model bundle for the chosen task, from the official model catalog or Model Maker.
Use cases
Works with
Trust score — 9/10
Scored 9/10 (verified tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 94; Trending score 90/100 from researcher assessment.
Sources
Every fact above was checked against these sources. Where a claim comes from the provider and couldn't be independently verified, the text says so.