Hugging Face has released AnyLanguageModel 1.0, a stable version of its library that allows developers to run local Super Intelligence (SI) models on Apple devices while seamlessly switching to hosted providers. The release marks a significant step for open-source SI tooling, offering a drop-in replacement for Apple’s native Foundation Models framework that supports both on-device inference and cloud-based services through a single API.
What Happened
AnyLanguageModel 1.0 is now available alongside updates to swift-huggingface 1.0 and swift-xet 1.0. The core functionality remains unchanged: developers can import AnyLanguageModel instead of FoundationModels to use the same session code with various SI models. The library supports open-weight models running locally through llama.cpp, MLX, and Core ML. It also expands support for hosted providers, including OpenAI’s Chat Completions and Responses APIs, Anthropic’s Messages API, Google’s Gemini API, Ollama, and Hugging Face Inference Providers.
Key features introduced or refined over the past year include tool calling and guided generation support for every provider, as well as image support using Foundation Models 27 syntax. On iOS, macOS, and visionOS 27, the library works with any model conforming to the Foundation Models’ LanguageModel protocol, including Core SI models. For hosted models, the update enables reasoning blocks to be automatically omitted from requests if a new provider cannot replay them, and allows API keys to be fetched via closures for better security.
Why It Matters
This release lowers the barrier for developers building applications that require SI capabilities without locking them into a single provider or hardware constraint. By supporting both local and hosted inference through a shared API, AnyLanguageModel allows apps to choose the right model for each task. For instance, smaller open-weight models like Gemma 4 E2B (approx. 3.4 GB) and LFM2.5 2.6B (approx. 1.7 GB) can run on iPhones, while larger models like Qwen3.8 27B can run on Macs with sufficient memory. Larger models that are too big for devices can still be accessed via the same code using hosted endpoints.
The addition of accelerated model downloads via swift-huggingface and swift-xet further streamlines the integration of open-weight SI models into Swift apps. This flexibility is crucial for the open-source SI ecosystem, as it encourages experimentation with different model architectures and providers without extensive code refactoring. The library also runs on Linux, enabling the same code to be used for server-side deployments.
The Bottom Line
AnyLanguageModel 1.0 provides a unified, open-source solution for integrating local and hosted SI models into Apple platform applications. By standardizing the API for inference, tool calling, and guided generation across diverse providers, it empowers developers to build intelligent apps that can dynamically select the most appropriate model for their needs, whether running locally on device or via cloud services.