Apple Core AI Framework: Core ML Replacement Coming at WWDC 2026

Apple is replacing Core ML with a brand-new framework called Core AI, set to debut at WWDC 2026 this June. The rename from "Machine Learning" to "AI" isn't cosmetic — it signals a fundamental shift in how developers integrate AI models into iOS and macOS apps. Third-party LLM support is the headline feature, and it could reshape how local AI runs on Apple hardware (Bloomberg/Mark Gurman, March 2026).

TL;DR: Apple will replace Core ML with Core AI at WWDC 2026 (iOS 27). The new framework lets developers plug third-party AI models — including LLMs — directly into apps. Combined with M5 Pro/Max hardware and the Foundation Models framework, Apple is building a complete local AI stack.

Why Is Apple Killing Core ML?

Core ML launched in 2017 to bring machine learning to iOS apps. It handled image classification, natural language processing, and on-device predictions. But the AI landscape has moved far beyond those use cases.

The term "machine learning" itself is now dated. LLMs, diffusion models, and agentic AI don't fit neatly into Core ML's original design. According to Bloomberg's Mark Gurman, Core AI will be a modernized replacement built for the current era of AI — one where developers need to run full language models, not just classifiers (9to5Mac, March 2026).

Core ML's limitations are well-known to developers:

• No native LLM support — running a language model required workarounds through llama.cpp or MLX
• Limited model formats — Core ML models (.mlmodel) didn't support modern transformer architectures well
• No third-party model ecosystem — every model had to be converted to Apple's format

Core AI aims to fix all three.

What Core AI Brings to the Table

The key shift: Core AI will let developers integrate outside AI models directly into their apps. That's a departure from Apple's historically closed approach to on-device AI.

Here's what we know from Gurman's reporting and related announcements:

The integration with Apple's Foundation Models framework — already available in macOS Tahoe for on-device intelligence — suggests a layered approach. Foundation Models handles Apple's own models (powering Apple Intelligence), while Core AI could serve as the bridge for third-party models like Llama, Qwen, or Mistral.

The Hardware Stack Is Already in Place

Core AI doesn't exist in isolation. Apple has spent the last year building hardware specifically designed for local AI workloads.

The M5 Pro and M5 Max, announced March 3, 2026, feature Neural Accelerators embedded in every GPU core — a first for any consumer chip. The M5 Max delivers up to 4x faster LLM prompt processing versus M4 Max, with 128GB unified memory and 614 GB/s bandwidth (Apple Newsroom, March 2026).

The timing is deliberate. Apple shipped the hardware in March. The software framework arrives in June. The full stack:

1. Silicon: M5 Pro/Max with 40 GPU Neural Accelerators + 16-core Neural Engine

2. Low-level framework: Foundation Models for Apple's own on-device models

3. Developer framework: Core AI for third-party model integration

4. User-facing AI: Siri with Gemini integration (expected iOS 26.5)

This is Apple building a complete local AI platform — not just a chip or a framework, but the entire pipeline from silicon to user experience.

What This Means for Ollama and MLX Users

If you're running local models on your Mac today, Core AI could change your workflow significantly.

The optimistic scenario: Core AI provides a native API for loading and running third-party LLMs. Developers of apps like LM Studio, Ollama, and MLX-based tools get a first-party framework to build on. This means better hardware utilization, tighter OS integration, and potentially faster inference than current community-driven solutions.

Apple already showed its hand by featuring LM Studio in official MacBook Pro press images — the first time Apple has publicly endorsed a third-party LLM tool.

The realistic concern: Apple might restrict Core AI to approved model formats or sizes. They could require models to pass a review process, similar to App Store approval. Or they might limit which Neural Engine features third-party models can access.

Either way, the existence of Core AI validates the thesis that local AI on Mac is a first-class use case, not a niche hobby. Apple is investing framework-level engineering resources into making it work.

MCP Integration: The Wildcard

One possibility Gurman flagged: Model Context Protocol (MCP) integration in Core AI. MCP is an open standard for connecting AI models to external tools and data sources. If Apple adopts it, any Core AI-powered app could connect to databases, APIs, file systems, and other services through a standardized protocol.

For local AI users, MCP support would mean your on-device LLM could access your files, calendar, email, and other Apple services — all running locally, with no cloud dependency. That's the privacy-first AI vision Apple has been selling since 2024.

MCP adoption by Apple would also accelerate the standard's growth industry-wide. As of early 2026, MCP has support from Anthropic, OpenAI, Google, and dozens of developer tools. Apple joining would make it the de facto integration protocol for AI applications.

The Siri Factor: Gemini + Core AI

Separately, Apple is expected to upgrade Siri with Google Gemini models in iOS 26.5 — a cloud-based enhancement. But Core AI serves a different purpose: enabling developers to build AI features that run entirely on-device (AppleInsider, March 2026).

This dual approach — cloud AI for Siri, local AI for developers — mirrors what the industry is settling on. Some tasks need massive models and internet access. Others (code completion, document summarization, image generation) work better locally, with no latency and no data leaving the device.

For the modelfit.io audience, Core AI is the piece that matters. It determines whether running a 14B or 30B model on your MacBook Pro becomes a system-supported capability rather than a community hack.

Timeline and What to Watch

Key questions for WWDC:

• Which model formats will Core AI support natively? (GGUF? MLX? Both?)
• Will third-party models get full Neural Accelerator access?
• Does Core AI include a model registry or marketplace?
• How does Core AI interact with Foundation Models?

We'll be covering WWDC 2026 in detail. If you're running local models on Apple Silicon, this is the most important developer conference in years.

---

FAQ

What is Apple Core AI?

Core AI is Apple's upcoming developer framework that replaces Core ML, expected to launch at WWDC 2026 with iOS 27. It enables third-party AI model integration in apps, supporting LLMs and modern AI architectures — a major upgrade from Core ML's classification-focused design.

Why is Apple replacing Core ML with Core AI?

Core ML was built in 2017 for traditional machine learning tasks like image classification. Modern AI requires LLM support, third-party model integration, and new hardware acceleration (Neural Accelerators in M5 chips). The rename from "ML" to "AI" reflects how the industry has moved beyond the "machine learning" label.

Will Core AI support Ollama and third-party LLMs?

According to Bloomberg's Mark Gurman, Core AI will let developers integrate outside AI models into their apps. While specific model format support (GGUF, MLX, safetensors) hasn't been confirmed, the framework's stated goal is third-party model compatibility — which would benefit tools like Ollama, LM Studio, and MLX.

When will Core AI be available?

Apple is expected to announce Core AI at WWDC in June 2026. It will ship with iOS 27 and the corresponding macOS release in fall 2026. Developers will likely get beta access immediately after the WWDC keynote.

Does Core AI work with the M5 Pro and M5 Max?

Yes. Core AI is designed to take full advantage of M5-series hardware, including the Neural Accelerators in every GPU core and the 16-core Neural Engine. The M5 Max with 128GB unified memory and 614 GB/s bandwidth is the ideal target hardware for running large local models through Core AI.