On This Page6 sections
FOUNDATION MODELS VS CORE ML / CORE AI VS MLX
Three of these are frameworks and one is a model, which is exactly why they get confused. The Foundation Models framework hands you Apple’s model through a session API. Core ML—and from WWDC26, Core AI—runs models you convert and ship. MLX is an array framework for training and running open models on Apple Silicon. FastVLM is an open model you can run through any of the last three.
| Feature | Foundation Models framework | Core ML → Core AI | MLX | FastVLM (model) |
|---|---|---|---|---|
| What it is | Swift/Python API over Apple’s built-in models (and, since WWDC26, any LanguageModel provider) | Runtime for models you convert; Core AI is the WWDC26 successor built for generative workloads | Open-source array framework (Python, Swift, C++) for Apple Silicon | Apple research VLM with open weights (0.5B–7B) |
| Model choice | Apple’s AFM 3 Core by default; PCC, MLX, Core AI or cloud providers via the protocol | Any model you convert from PyTorch | Any open model with an MLX implementation (mlx-lm, mlx-vlm) | Fixed: FastVLM-0.5B / 1.5B / 7B |
| What you ship | Nothing—the model is part of the OS | Compiled model package inside your app | Weights (downloaded or bundled) plus the MLX runtime | Weights via the runtime you choose |
| Runs on | Apple Intelligence devices; macOS for Python/CLI | iPhone, iPad, Mac, Vision Pro, Apple TV, Watch | Apple Silicon Macs; iOS through MLX Swift | Browser (WebGPU), Python/CUDA, Mac (MLX), iPhone (MLX/Core ML) |
| Vision input | Images since WWDC26; Vision framework tools callable by the model | Whatever the converted model supports | Whatever the model supports (mlx-vlm covers FastVLM) | Yes—its whole purpose |
| Best for | Fast integration, structured output, no weights to manage | Custom or fine-tuned models with full control and ANE/GPU acceleration | Research, fine-tuning and running open LLMs/VLMs on a Mac | High-resolution image Q&A with low first-token latency |
Checked 2026-09-17
Decision summary
- Start with the Foundation Models framework if you only need Apple-platform apps and the built-in model is good enough.
- Reach for Core AI (or Core ML on older toolchains) when you must ship a specific converted model with full control.
- Use MLX for experiments and open models on a Mac; FastVLM has an official MLX app.
- FastVLM is not an alternative to these frameworks—it is a model that runs through them, or in the browser.
Pick the framework by constraint
- No model management, Apple-only → Foundation Models framework
- Custom model, on-device, full control → Core AI / Core ML
- Mac research, fine-tuning, open weights → MLX
- Web or Linux included → open model such as FastVLM via Transformers or Transformers.js
How to read this table
Core AI was announced at WWDC26 as the successor to Core ML; availability depends on the OS and Xcode versions you target. Check Apple’s documentation for the exact deployment targets before choosing.
Questions
Can I run FastVLM through the Foundation Models framework?
Since WWDC26 the framework accepts any provider that conforms to LanguageModel, including MLXLanguageModel. Whether a vision model like FastVLM is exposed that way depends on the provider implementation; the documented FastVLM paths remain MLX, Core ML, Transformers and the browser.
Is Core ML deprecated?
Apple positions Core AI as the successor for new generative workloads, but existing Core ML models and APIs continue to work on the platforms they shipped for. Plan new projects on Core AI and migrate existing ones on your own schedule.
Sources
Continue
Continue Reading
More articles connected to the same themes, protocols, and tools.
Referenced Tools
Browse entries that are adjacent to the topics covered in this article.
