AnyLanguageModel 1.0: One Year of Local Models on Apple Platforms

AnyLanguageModel 1.0 launches as a drop-in replacement for Apple’s Foundation Models framework, enabling developers to select models and providers without rewriting session code. The update supports local inference via llama.cpp, MLX, and Core ML, as well as hosted providers, all through a shared API. On OS 27, it also works with models conforming to the Foundation Models’ LanguageModel protocol, including Core AI models. The framework retains its original goal of reducing complexity for app developers integrating language models.

Swift-huggingface 1.0 and swift-xet 1.0 accompany the release, accelerating model downloads from the Hugging Face Hub directly into Swift applications. Inference providers now support open-weight models too large for local devices, offering guided generation, tool calling with streaming, and token usage tracking. Reasoning blocks are automatically excluded when switching providers unable to replay them, ensuring compatibility across different services.

API keys are now managed via closures, allowing apps to fetch short-lived tokens from a server instead of embedding static strings. The README provides guidance on when to use user-provided keys versus proxy routing. AnyLanguageModel’s compatibility extends to Linux, enabling server-side deployment with the same codebase used in Apple platforms. The update reflects a year of community-driven improvements and model advancements.

Local models have grown more capable, with popular options like Qwen3.8 27B requiring 32 GB of Mac memory and GLM-5.3 from Z.ai needing 1.5 TB in BF16 for local deployment. Developers can select MLX models with a single line of code, while larger models remain accessible through inference providers. Migration guidance is available for upgrading existing apps using AnyLanguageModel, swift-huggingface, and swift-xet.