Google has open-sourced ML Drift under Apache 2.0, the GPU acceleration engine beneath its LiteRT runtime for on-device inference. The pitch is portability: one layer that talks to whatever graphics hardware a phone, laptop or browser happens to offer.
Developers currently pick their way through Metal on Apple hardware, OpenCL on Android, OpenGL ES on older devices and WebGPU in browsers. ML Drift absorbs those differences so the same model runs across them, from computer vision and audio through to generative workloads.
Underneath, it adds unified shaders through tensor virtualization, extensible custom operators and support for five-dimensional tensors. Google also tuned for GPUs from Arm, Intel and Qualcomm.
Desktop inference gets a path too. On Windows and Linux it reaches hardware through Dawn, Chromium’s WebGPU implementation, which avoids writing separate DirectX and Vulkan paths. Metal remains native on macOS.
Google benchmarked with Gemma models on workstation-class machines and reports 12 percent lower memory overhead than other frameworks, including llama.cpp and MLX LM. It also points out that the engine is already in production at scale, powering segmentation effects in YouTube Shorts, computational photography in Photos, and calling in Meet.
Published gains range from up to 40 percent lower frame latency on Android and iOS segmentation to roughly two seconds saved in photo pipelines and 30 percent faster edits in Adobe Lightroom and Photoshop.
Open-sourcing it lets teams build custom runtimes on the same foundation rather than wait for a vendor path.
