Docker 24.1 Optimizes Images for Edge AI

Docker announced version 24.1 with improvements centered on optimizing container images for deployment on edge devices with limited resources — a segment that has been growing rapidly with the expansion of AI inference workloads in industrial, healthcare, and urban infrastructure environments. The update includes a new optimized multi-stage builder, automatic layer reduction, and native support for image formats specific to ARM and RISC-V architectures.

What Changed for AI Images on Edge

The challenge of distributing AI models on edge devices is twofold: container images that include frameworks like PyTorch, TensorFlow Lite, or ONNX Runtime tend to be large — frequently above 2 GB —, and edge devices have limited bandwidth, restricted storage, and often no permanent internet connection for updates. Docker 24.1 attacks this problem with a Layer Deduplication Engine that identifies redundant layers between related images and stores only a delta per version.

In tests with a typical inference stack based on ONNX Runtime with ARM integrated GPU support, the size reduction ranged between 34% and 61% depending on the base image used. The Builder also began emitting warnings when it detects dependencies included in the final image that are only needed at build time, guiding the developer to move them to temporary stages.

RISC-V Support and Ecosystem Impact

Support for RISC-V architecture images, still experimental but functional, is a signal of the growing ecosystem in industrial hardware and next-generation IoT devices. RISC-V chips with neural inference accelerators are reaching the market at much lower prices than ARM alternatives, and Docker compatibility facilitates the portability of already containerized workloads to this platform. Version 24.1 is available as an automatic update for Docker Desktop users and as a package in official repositories for Linux distributions.