Google open-sources ML Drift, the GPU engine inside LiteRT, and stops adding features to the TFLite GPU delegate

On October 8, 2026, Google's AI Edge team published ML Drift as open source under the Apache 2.0 license. ML Drift is a GPU compute engine for on-device inference: it runs machine learning models on the graphics processor of a phone, a browser, or a laptop instead of on a server. It already powers GPU acceleration in LiteRT, Google's runtime for on-device models, and the same code is now available as a standalone C++ library.

What happens to the TFLite GPU delegate

With this release, the legacy TFLite GPU delegate will no longer receive new feature updates. The delegate is the component that has handed TensorFlow Lite models to the GPU until now. Google asks developers to move to the LiteRT ML Drift GPU accelerator instead.

Google says the new accelerator is fully backward compatible with existing models, and that models already running on the old delegate get faster right away. Those are the vendor's own claims, and the post cites no outside test of the migration.

On Android, ML Drift acceleration is available now in the standalone LiteRT packages. LiteRT delivered through Google Play Services gets it later, and Google gives no date for that.

This suggests that an app on the old delegate keeps working, but that new GPU capabilities, like the ones described below, will reach developers only through LiteRT.

What the engine changes

The central change is what Google calls tensor virtualization. A tensor is the multidimensional array of numbers a model computes on. The old delegate tied each tensor to a specific GPU object, such as a texture or a buffer, separately for OpenGL, OpenCL, and Metal, so each backend needed its own shader code. ML Drift resolves that mapping when the model is initialized, so one set of shader templates serves every backend.

The engine also supports 5D tensors. The old delegate was limited to 4D, which forced workarounds for 3D convolutional networks and for models that process space and time together. Google names YOLO 11n, MobileViT v2, and Swin Transformer v2 as models that now run directly on the GPU.

For large language models, ML Drift switches kernels and memory layouts between the two phases of generation: prefill, when the model reads the whole prompt, and decode, when it produces tokens one at a time. It also ships a SKILL.md guide written for coding agents that author, register, and verify custom shaders.

Where it runs and what Google measured

ML Drift supports OpenGL ES, OpenCL, Metal, and WebGPU, which covers Android, iOS, and the browser. A desktop preview runs on Windows and Linux through Dawn, Chromium's WebGPU implementation, which lets the same WebGPU code run natively outside the browser. On macOS it runs through Metal. Chrome already uses ML Drift to run Gemini Nano behind its built-in AI APIs, including the Prompt, Summarize, and Writer APIs.

Every performance figure in the announcement comes from Google or its partners. Google says YouTube Shorts effects saw up to 40% lower average frame latency on Android and iOS, and Adobe reports up to 30% faster on-device performance for Select Subject, Select Sky, and Adaptive Portrait in its mobile apps. In Google's Gemma tests, ML Drift had up to 12% lower memory overhead than other frameworks, which the post does not name. No independent measurement has been published, and the desktop results appear only as charts.