Google open-sources ML Drift on-device GPU engine

On 2026-11-05, InfoQ reported that Google has open-sourced ML Drift, a cross-platform GPU compute engine for on-device AI inference, released under the Apache 2.0 license. ML Drift abstracts away the complexity of different GPU hardware and low-level APIs, unifying OpenGL ES, OpenCL, Metal and WebGPU under a single layer. It serves as the GPU acceleration backend that powers LiteRT, letting developers build custom graphics rendering and inference runtimes for applications ranging from traditional computer vision, audio and depth processing to generative AI.

Google says ML Drift aims to establish a new foundation for portability and performance, combining the stability of traditional ML with the peak performance needed for frontier generative AI. Compared with the older TFLite GPU delegate, ML Drift introduces a unified shader architecture built on tensor virtualization, extensible custom operators and native support for five-dimensional tensors. It also carries hardware-specific optimizations for GPUs from Arm, Intel and Qualcomm. Google claims that on the Gemma model its memory overhead is 12% lower than llama.cpp and MLX LM, and that it is already deployed in Chrome, YouTube Shorts, Photos and Meet. The report also notes that ML Drift is built on Dawn, Chromium's WebGPU implementation. No further developments have been reported since the release, so the current status of the project is best tracked through Google's official repository.