ML Drift: Next-Gen GPU AI/ML Inference at the Edge
This is a technical, developer-facing release, relevant mainly if your organisation builds its own on-device AI apps for mobile or edge devices and needs faster, more memory-efficient GPU inference. It has no direct impact on how you use consumer AI products like Gemini.
What changed
Source summary
Google's AI Edge team open-sourced ML Drift, a cross-platform GPU compute engine for on-device AI inference that runs on OpenGL ES, OpenCL, Metal, and WebGPU, serving as the GPU acceleration engine inside LiteRT and usable as a standalone library. It includes a tensor virtualisation layer so one shader codebase works across GPU backends, a custom-ops registration API with an agent-oriented SKILL.md guide for coding agents, 5D tensor support for models like 3D convolutional networks, LLM-specific kernel/layout optimisations for prefill and decode stages, and early desktop previews (WebGPU on Windows/Linux, Metal on macOS). In Gemma benchmarks Google reports up to 12% less memory use than other frameworks. The code is open source under Apache 2.0; standalone LiteRT packages with ML Drift are available now for Android developers using unbundled runtimes, with support via Google Play Services described as "coming soon". The legacy TFLite GPU delegate will no longer receive new features. Desktop support and benchmarks are described as an early preview/snapshot, with mobile and edge devices the primary focus.
Why it matters to your work
Practical interpretation
This is a technical, developer-facing release, relevant mainly if your organisation builds its own on-device AI apps for mobile or edge devices and needs faster, more memory-efficient GPU inference. It has no direct impact on how you use consumer AI products like Gemini.
Source & availability
This is a summary of a published provider announcement, with practical context from SCALA’s feed.
Plan, region, workspace and rollout conditions may affect access in Australia. Check the original source for current availability.
developers.googleblog.comThis update is primarily for technical teams.