Apple's Foundation Models framework gives iPhone and Mac developers a clean way to call the language model built into the operating system, but only that model. Hugging Face's AnyLanguageModel, now at version 1.0 after about a year of releases, keeps Apple's API and lets the same code talk to almost any model instead. For developers curious about local models but wary of supporting several inference providers, it lowers the cost of finding out.
The pitch is a one-line change. Swap the import, and the same session code runs against Apple's system model, open-weight models through llama.cpp, MLX or Core ML, or hosted services, including Hugging Face Inference Providers and OpenAI-compatible endpoints such as OpenRouter. Version 1.0 brings image support using the Foundation Models 27 syntax, so prompts can include pictures, and on Apple's 27 releases it also works with any model that follows Apple's LanguageModel protocol. Hosted providers get the same features as local ones: guided generation, tool calling with streaming, and token counts. Reasoning is kept in the transcript and dropped automatically when a request moves to a provider that cannot replay it, and API keys come from a function rather than a fixed string, so an app can fetch short-lived tokens from its own server.
Hugging Face has also released swift-huggingface 1.0 and swift-xet 1.0 to speed up model downloads from the Hub. Its examples show the range: some of the most downloaded small models fit on an iPhone, Qwen3.8 27B runs on a Mac with 32 GB of memory or more, and something the size of GLM-5.3, about 1.5 TB in BF16, is better reached through hosted providers.
For anyone building on Apple's platforms, this makes local-first apps more practical. It sits on top of the runtimes we review, MLX and llama.cpp, rather than replacing them, and it means a developer can start with Apple's built-in model and move to an open or hosted one later without rewriting the app.
The main risk is dependence on Apple's roadmap. AnyLanguageModel mirrors an API Apple changes every year, so each OS release creates work to keep pace, and features Apple adds may arrive here late. Developers also take on the job of choosing and testing models themselves, including how they handle privacy when requests leave the device. Hugging Face says models that make decisions and models that listen are coming next, which would widen the scope beyond text.