> ## Content Index
> Fetch the complete content index at: https://globalfeed.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# Google open-sources ML Drift, its GPU engine for on-device AI
- URL: https://globalfeed.ai/en/google-open-sources-ml-drift-its-gpu-engine-for-on-device-ai/
- Published: 2026-10-10T18:55:19.000Z
- Updated: 2026-10-10T18:55:19.000Z
- Description: Google released ML Drift, the engine that runs AI models on the GPU of phones, browsers and computers, as open source under the Apache 2.0 license. It replaces the old TensorFlow Lite GPU delegate and already powers YouTube Shorts, Google Photos, Chrome and Meet.
- Author: GlobalFeed Editor
- Tags: Google, ML Drift, LiteRT, on-device ai, x-googledevs, dil-en, elle

**Google** released **ML Drift** as open source under the Apache 2.0 license on October 8, 2026\. ML Drift is an inference engine that runs AI models on a device's own graphics chip (GPU). It is the layer that lets a model run fast on a phone, in a browser or on a computer without sending data to the internet. According to Google's announcement, ML Drift is the direct successor to the old TensorFlow Lite GPU delegate, which will no longer receive new features. The engine works as the GPU acceleration core of LiteRT, Google's on-device AI framework, and is also available as a standalone library.

According to the announcement, ML Drift already runs in YouTube Shorts, Google Photos, Meet and the Gemini Nano based built-in AI APIs in Chrome. Google AI Edge's Gallery and Eloquent apps, and Foresight, its offline meeting notes app for Mac, also use it.

## Which devices and which models

ML Drift runs on OpenGL ES, OpenCL, Metal and WebGPU. Desktop support is an early preview: it runs through WebGPU (via Dawn) on Windows and Linux and through Metal on macOS. Besides vision, audio and depth models, the engine supports large language models and large diffusion models, with separate optimizations for the stage where the prompt is processed and the stage where the answer is generated.

In Google's measurements, ML Drift cut average frame latency in YouTube Shorts by up to 40 percent and made Google Photos up to 2 seconds faster than the old GPU delegate. Google also says it uses up to 12 percent less memory than other frameworks in its Gemma benchmarks. According to partner statements in the announcement, Adobe saw up to 30 percent faster performance for features such as Select Subject and Select Sky in Lightroom and Photoshop, and Snap saw 30 percent lower latency for face and style effects in Snapchat. Arm, Intel and Qualcomm are among the hardware partners contributing for their own GPUs.

## How developers can use it

App developers can use the engine through LiteRT's ML Drift GPU accelerator. For GPU and systems engineers, the GitHub repository offers C++ APIs, support for custom shaders and quickstarts for OpenCL on Android and for WebGPU. ML Drift already ships in the standalone LiteRT packages on Android; delivery through Google Play Services is coming soon.

Desktop support is still an early preview, and the engine's main focus remains mobile and edge devices. The speed and memory figures were measured by Google and its partners. For a fuller picture, follow evaluations from independent researchers as well.

Sources: [Google for Developers: ML Drift announcement](https://developers.googleblog.com/ml-drift-next-gen-gpu-aiml-inference-at-the-edge/?ref=globalfeed.ai), [GitHub: ML Drift](https://github.com/google-ai-edge/ml-drift?ref=globalfeed.ai), [LiteRT: ML Drift GPU accelerator](https://developers.google.com/edge/litert/next/gpu?ref=globalfeed.ai).