Hugging Face released AnyLanguageModel 1.0 on October 20, 2026, marking a year of development on its replacement framework for Apple's Foundation Models API. The release shipped alongside version 1.0 releases of swift-huggingface and swift-xet, which handle model downloads from the Hugging Face Hub to Swift applications.
Developers switch to the framework by replacing FoundationModels with AnyLanguageModel in import statements. That change allows identical session code to run on Apple's system model, local open-weight models executed through llama.cpp, MLX, or Core ML, and external hosted endpoints. The library also runs on Linux servers.
On OS 27, AnyLanguageModel works with any model that conforms to the Foundation Models LanguageModel protocol, including Core AI models. The framework adds image prompt support aligned with Foundation Models 27 syntax on iOS, macOS, and visionOS 27. Tool calling and guided generation, including nested @Generable types and arrays, function across every supported provider.
The software routes hosted requests to OpenAI's Chat Completions and Responses APIs, Anthropic's Messages API with adaptive thinking and effort settings, and Private Cloud Compute on OS 27. It also connects to Hugging Face Inference Providers and OpenRouter through custom base URLs. Reasoning blocks are automatically stripped from transcripts if a target provider cannot replay them, and API keys can be supplied via closures to pull short-lived tokens from private servers.
Device-level model targets highlighted by Hugging Face include Google's 3.4 GB Gemma 4 E2B and Liquid AI's 1.7 GB LFM2.5 2.6B, both of which run on iPhones in GGUF formats. Alibaba's 16 GB Qwen3.8 27B runs on Macs equipped with 32 GB of memory or more, while Z.ai's 1.5 TB GLM-5.3 requires four Mac Studios or hosted routing. Hugging Face stated it plans to expand the project to models that decide and models that listen.
