Apple announces CoreML-Edge v2 optimized for on-device generative models

Tech · 4 min read

CoreML-Edge v2 introduces quantization-aware runtimes, faster tensor fusion for M-series and A-series chips, and an on-device tokenizer optimized for low-memory scenarios. Apple highlighted new APIs that let apps stream partial generative outputs and gracefully recover from interruptions on mobile.

The update specifically addresses creative workflows: designers building on-device sketch-to-design apps or AR prototyping tools can now run larger model families without offloading to servers. Apple also released sample models and reference code for image-to-layout and microcopy generation.

Privacy and security are emphasized: CoreML-Edge v2 includes enhanced entitlements and sandboxed model execution to prevent exfiltration of user data. Apple expects the improvements to encourage a wave of apps focused on offline creative assistance and secure sharing.