By Apple killed Core ML today. After nine years as the default on-device ML framework for iOS and macOS developers, it is being replaced by Core AI — announced at this morning’s WWDC 2026 keynote. If you have .mlmodel files in your Xcode projects, here is what you need to know, and how worried you should actually be.
It’s Not a Rebrand. The Architecture Changed.
Core ML was built in 2017 to run image classifiers, text regressors, and tree ensembles on the Neural Engine. That is what existed then. It did the job well for a decade, but it was never designed for large language models, streaming token generation, or agent-style tool calling — the things developers are building now.
Core AI is Apple’s answer to that gap. It handles async inference natively, supports large model memory footprints, and opens the door to third-party model integration. Developers will be able to plug in their own models — a fine-tuned Llama variant, a custom Mistral deployment — without being locked to Apple’s .mlmodel format. That is a significant departure from how Core ML worked.
The name change from “ML” to “AI” is not marketing spin. It is Apple acknowledging that the framework built for 2017’s machine learning problems cannot carry 2026’s generative AI requirements.
Your Existing Core ML Apps Are Fine
Before you start a migration panic: Core ML and Core AI will coexist in iOS 27. Apple has not announced a deprecation date. Your existing .mlmodel files and MLModel API calls will keep working.
The urgency is not about what breaks — nothing breaks on day one. The urgency is about what you are locked out of. Developers building new AI features in iOS 27 apps — Siri Extensions, Writing Tools integration, Foundation Models access — will need Core AI. Core ML does not connect to any of those new APIs. If you are adding generative features to an existing app, you are writing Core AI code from the start.
The MCP Piece Is the Real Story
Core AI reportedly includes Model Context Protocol support. If that holds up in the official documentation, it is the most significant part of this announcement.
MCP is the open standard developed by Anthropic that lets AI models act as agents: call tools, read and write data, trigger workflows inside applications. The difference between “AI that answers questions” and “AI that actually does things.” Microsoft adopted it. Google followed. If Apple ships MCP natively in Core AI, any model plugged into the framework can interact with your app’s data and capabilities through a standardized interface.
For developers, this means two things. First, your app can expose its capabilities as MCP tools — making it accessible to any AI model running through Core AI. Second, you can build apps that consume AI capabilities through MCP without writing custom integration code for every model. One protocol, any model.
Foundation Models: Now Unified
iOS 26 shipped the Foundation Models framework — Apple’s on-device LLM, available to all apps without an API key. It has been running separately from Core ML since last year, which created a split architecture that developers had to work around.
Core AI unifies them. The Foundation Models LLM, third-party models, and custom model deployments all sit under one API surface. The system handles dispatch — deciding whether inference runs on-device or delegates to cloud based on model size, task requirements, and user privacy settings. On-device remains the default. Cloud delegation requires explicit user permission.
What to Do Right Now
- Install Xcode 18 beta — ships with the iOS 27 SDK and Core AI APIs. Get familiar with the new framework before it ships to users in the fall.
- Watch the WWDC sessions — “What’s New in Core AI” and the Core ML migration session will have the authoritative migration path. Do not guess at API changes before watching them.
- Audit your Core ML usage — identify which models in your app are doing classification or regression tasks that could eventually be replaced by Foundation Models API calls. Not urgent, but worth mapping now.
- New AI features: use Core AI — if you are adding anything LLM-related to an iOS 27 app, start with Core AI. Core ML does not give you access to the new stack.
Core AI’s real significance is not that it makes existing apps better — it is that it gives developers the infrastructure to build on-device AI agents that actually do things. Apple just made that infrastructure part of the OS. 9to5Mac’s original March report on Core AI has proved accurate on every detail confirmed so far. What you build on top of it is the question worth spending the rest of WWDC week thinking about.
