# Google open-sources ML Drift, a cross-platform GPU engine for on-device inference

published: 2026-10-25T17:05:28.795Z
topic: infrastructure
source: https://www.infoq.com/news/2026/10/google-ml-drift/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
url: https://aiagentnews.io/stories/google-open-sources-ml-drift-a-cross-platform-gpu-engine-for-on-device-inference

Google has released ML Drift under Apache 2.0, a cross-platform GPU compute engine for on-device AI and machine learning inference. It is the core GPU acceleration engine behind LiteRT, and Google will stop updating the legacy TensorFlow Lite GPU delegate.

ML Drift hides differences between GPU architectures, drivers and low-level APIs, with support for OpenGL ES, OpenCL, Metal and WebGPU. Compared with the legacy TFLite GPU delegate, it adds unified shaders through tensor virtualization, extensible custom operations and five-dimensional tensor support. It also includes hardware-specific optimizations for GPUs from Arm, Intel and Qualcomm Technologies.

On desktop, ML Drift runs inference on Windows and Linux through Dawn, Chromium's WebGPU implementation, and keeps native Metal support on macOS. In Google's benchmarks using Gemma models on developer-grade workstation hardware, it reported 12% lower memory overhead than llama.cpp and MLX LM. Google says the engine already runs on millions of devices through Chrome, YouTube Shorts, Photos and Meet.

Google cited gains in individual features, including up to 40% lower average frame latency for segmentation effects in YouTube Shorts and up to 30% faster selected AI photo-editing features in Adobe Lightroom and Photoshop. Snapchat reported 30% lower model latency for face and style effects on Android. Vladislav Busel, founder of Tensorust, asked how ML Drift compares with Core ML on recent iPhones, and Google has not published comparative figures.

## Why it matters if you build agents

Developers targeting edge GPUs can use the LiteRT ML Drift accelerator directly, or the lower-level C++ API to build custom compute graphs and shaders. Teams still relying on the TFLite GPU delegate should plan a migration to LiteRT ML Drift, since Google will no longer update the legacy delegate. The benchmarks are Google's own, so measure latency and memory on your target devices before relying on the gains.