On 8 October 2026 Google said it has open-sourced ML Drift, the engine that runs AI models on a device’s graphics chip (GPU). The release is under the Apache 2.0 licence, a permissive open-source licence that allows commercial use [1][2]. The speed and memory figures below are Google’s own claims, not independent tests.
Spotted via @ByteVireo on X
What it is
Google’s Developers Blog calls ML Drift “our high-performance, cross-platform, on-device GPU compute engine specifically built for on-device AI/ML inference” [1]. Inference means running an already-trained model to get answers. The blog says it works across the graphics interfaces OpenGL ES, OpenCL, Metal and WebGPU, and that ML Drift is “the core GPU acceleration engine within LiteRT”, Google’s runtime for running AI models on devices, and also “available as a standalone library” [1]. Our earlier piece on EmbeddingGemma 2 lists LiteRT among its supported tools.
Google’s figures
All three come from Google’s blog [1]:
- YouTube Shorts: “up to a 40% reduction in average frame latency” on Android and iOS for its segmentation-based effects. Latency is the delay per video frame.
- Adobe: Lightroom and Photoshop features including Select Subject, Select Sky and Adaptive Portrait were “up to 30% faster” on-device.
- Memory: in Google’s Gemma benchmarks, “up to 12% lower memory overhead than other frameworks”. The text does not name the frameworks.
“Up to” means best cases, not typical results.
What changes for developers
The blog says the older TFLite GPU delegate, a plug-in that hands work to the GPU, “will no longer receive new feature updates”, and urges developers to move to the LiteRT ML Drift accelerator. Standalone LiteRT packages have it now; LiteRT in Google Play Services is “coming soon” [1].
Checked on GitHub
The repository google-ai-edge/ml-drift is public, and GitHub lists its licence as Apache 2.0 [2].
What this does not show
- Independent speed tests. No third party has confirmed the figures.
- Typical gains. The numbers are “up to” and tied to specific features and models.
- Which rivals were beaten. The memory comparison names no framework.
- A date for Play Services. Google gives none.
Related on TSN: EmbeddingGemma 2: Google’s open multimodal embeddings for the edge
Sources
- Chintan Parikh and Juhyun Lee, “ML Drift: Next-Gen GPU AI/ML Inference at the Edge”, Google Developers Blog, 8 October 2026 (read in full; figures are Google’s own). https://developers.googleblog.com/en/ml-drift-next-gen-gpu-aiml-inference-at-the-edge/
- google-ai-edge, ml-drift repository and LICENSE (Apache License 2.0), GitHub (checked 10 October 2026). https://github.com/google-ai-edge/ml-drift
- @ByteVireo, post on X, 9 October 2026 (aggregator repost; credit only). https://x.com/ByteVireo/status/2108532825162949102
