AnyLanguageModel 1.0: One Year of Local Models on Apple Platforms

Hugging Face says AnyLanguageModel has reached 1.0 after a year of releases, and the pitch is still the same: use one Foundation Models-style API, but pick the model and provider that actually fit the job. The library started as a drop-in replacement for Apple’s Foundation Models framework, aimed at developers who liked local models but didn’t want to stitch together a pile of inference backends by hand.

The basic swap is still tiny. Change the import from FoundationModels to AnyLanguageModel, and the same session code can talk to Apple’s system model, open-weight models through llama.cpp, MLX, and Core ML, or a range of hosted providers. Over the past year, the project added tool calling everywhere, guided generation everywhere, nested @Generable types and arrays, plus image support using Foundation Models 27 syntax on iOS, macOS, and visionOS 27.

The hosted side grew in parallel. AnyLanguageModel now supports OpenAI’s Chat Completions and Responses APIs, Anthropic’s Messages API with adaptive thinking and effort options, Google’s Gemini API, Ollama Private Cloud Compute on OS 27, and Hugging Face Inference Providers through a custom base URL. That last bit matters because it lets developers use big open-weight models that won’t fit on a device while keeping the same session code. Hosted models also get guided generation, streaming tool calling, token usage, and reasoning blocks that are dropped automatically when a new provider can’t replay them.

Hugging Face also pushed the plumbing forward. API keys now come from a closure instead of a string, so apps can fetch short-lived tokens from their own server, and the README points developers toward either bring-your-own-key setups or proxying requests. The package runs on Linux too, which turns the same code into something that can live on a server, not just on Apple hardware.

The model side has moved just as quickly. Hugging Face highlights Gemma 4 E2B from Google at about 3.4 GB in GGUF, LFM2.5 2.6B from Liquid AI at about 1.7 GB, and Qwen3.8 27B for MLX at about 16 GB. The first two fit on an iPhone, while the Qwen model needs a Mac with 32 GB of memory or more. For people with very large Macs lying around, GLM-5.3 from Z.ai is also available, though not exactly in a casual way: it takes about 1.5 TB in BF16.

The release wraps with the usual 1.0 upgrade path and migration notes for AnyLanguageModel, swift-huggingface, and swift-xet. But the bigger signal is simple: Hugging Face is betting that app developers want one Swift surface for local, hosted, and server-side model work, and it’s building the whole stack around that idea.