AnyLanguageModel 1.0: One Year of Local Models on Apple Platforms

Introduction

One year after its debut, AnyLanguageModel 1.0 reaches a milestone: developers can now run a wide range of language models—both locally on iOS, macOS, and visionOS devices and remotely via hosted providers—through a single, drop‑in compatible API. The release bundles updated Swift packages (`swift‑huggingface` and `swift‑xet`), richer feature sets such as tool calling, guided generation, and image‑augmented prompts, and tighter integration with Apple’s Foundation Models framework.

A unified API for every inference scenario

The core promise of AnyLanguageModel remains unchanged: write once, run anywhere. By swapping a single import statement (`import FoundationModels` → `import AnyLanguageModel`) the same session code works with:

• System‑provided models on Apple silicon.
• Open‑weight models compiled with llama.cpp, MLX, or Core ML.
• Hosted endpoints from OpenAI, Anthropic, Google Gemini, Ollama, Hugging Face Inference, and any OpenAI‑compatible service.

Key API enhancements include:

• Tool calling and guided generation that operate uniformly across providers, even for nested `@Generable` types.
• Image support using the Foundation Models 27 syntax, enabling multimodal prompts on iOS, macOS, and visionOS.
• Automatic handling of reasoning blocks—they are omitted when a provider cannot replay them, preserving transcript fidelity.
• API‑key retrieval via closures, allowing short‑lived tokens to be fetched securely from a backend.

Swift libraries that accelerate model delivery

Two companion packages ship at version 1.0:

• swift‑huggingface – streams model files from the Hugging Face Hub directly into Swift apps, handling caching and quantization automatically.
• swift‑xet – provides accelerated download and unpacking for large models hosted on external storage.

Both libraries expose a concise initializer, e.g.: ```swift let model = MLXLanguageModel(modelId: "mlx-community/Qwen3.8-27B-4bit") ``` Developers can now pull a 27‑billion‑parameter model onto a Mac with 32 GB + RAM in a single line of code.

Local models become more capable

The model ecosystem has matured dramatically. Popular open‑weight models now fit on consumer Apple devices:

• Gemma‑4 E2B (3.4 GB, Q4_0 quantized) – runs comfortably on an iPhone.
• LFM2.5‑2.6B (1.7 GB, Q4KM) – a lightweight alternative for on‑device chat.
• Qwen3.8‑27B‑4bit (≈16 GB) – feasible on high‑end Macs.

For ultra‑large models like GLM‑5.3 (≈1.5 TB), the framework falls back to hosted inference, preserving a consistent developer experience.

Migration path to 1.0

Existing projects can upgrade by adjusting their Swift Package Manager entry: ```swift .package(url: "https://github.com/huggingface/AnyLanguageModel", from: "1.0.0") ``` Release notes for AnyLanguageModel, swift‑huggingface, and swift‑xet detail migration steps, including trait flags for MLX support and updated documentation on token handling.

Looking ahead

The team signals two new model families in the pipeline:

• Decision models that output structured actions rather than text.
• Listening models for audio‑first interactions.

Both will expose the same Swift‑friendly API, reinforcing the vision of a single entry point for all AI capabilities on Apple platforms.

--- Practical take‑away: If you’re building an iOS or macOS app that needs language‑model intelligence, adopt AnyLanguageModel 1.0 to simplify provider selection, leverage the latest open‑weight models locally, and future‑proof your code for upcoming decision‑ and audio‑centric models.