AnyLanguageModel 1.0 brings Apple's on-device and hosted models under one Swift API

Hugging Face published AnyLanguageModel 1.0 on October 20, alongside 1.0 releases of swift-huggingface and swift-xet. The Swift library first appeared last November as a drop-in replacement for Apple's Foundation Models framework, built so developers could keep one API and still choose which model or provider their app talks to.

The upgrade path is deliberately small: developers swap import FoundationModels for import AnyLanguageModel, and existing session code keeps working. The same session can now point at the on-device system model, at open-weight models running through llama.cpp, MLX or Core ML, or at a hosted provider. Local and hosted inference stop being an either-or choice.

Features added over the past year sit in three places. Tool calling works with every provider, and guided generation covers all of them too, including nested @Generable types and arrays of them. Image support follows Foundation Models 27 syntax, including with the system model on iOS, macOS and visionOS 27. On OS 27, any model that conforms to the LanguageModel protocol can be plugged in, Core AI models included.

Hosted support now covers OpenAI's Chat Completions and Responses APIs, Open Responses, Anthropic's Messages API with adaptive thinking and effort options, Google's Gemini API, Ollama, Private Cloud Compute on OS 27, and OpenAI-compatible endpoints such as Hugging Face Inference Providers and OpenRouter through a custom base URL.

Hosted providers inherit the same features as local ones: guided generation, streamed tool calling, and token usage on responses. Reasoning is now part of the transcript, and blocks are dropped automatically when a new provider cannot replay them. API keys come from a closure rather than a hard-coded string, so apps can fetch short-lived tokens, and the library also runs on Linux so server-side code can share the same session logic.

The blog's open-weight examples show how far small models have come. Gemma 4 E2B ships in GGUF at roughly 3.4 GB with Q4_0 quantization, LFM2.5 2.6B at about 1.7 GB, both small enough for an iPhone; Qwen3.8 27B for MLX needs about 16 GB and a Mac with 32 GB of memory; GLM-5.3 from Z.ai would take roughly 1.5 TB in BF16, so most teams will reach it through Inference Providers.

Upgrading means pointing a Swift Package dependency at 1.0.0 and opting into the MLX trait, with migration notes published for all three projects. The post closes by saying models that decide and models that listen are coming next, a hint that the same API is meant to cover more than text.

For Apple developers the significance is not any single feature but the collapse of model choice into a parameter: one session object can sit in front of a small on-device model or a hosted model far too large for the device. That is what makes provider strategy something an app can change late.