AnyLanguageModel 1.0 Simplifies Local and Hosted Model Choice for Apple Developers
Swift developers can swap one import to access open-weight models on-device via MLX, llama.cpp or Core ML, or hosted providers, all with the same Foundation Models-style…
- A single import change lets existing Swift session code work with local quantized models that fit on iPhones or Macs.
- Tool calling, guided generation for nested types, and image prompt support now work uniformly across every local and hosted provider.
- Hosted open-weight models too large for devices can be used through Hugging Face Inference Providers or OpenRouter with identical code.
What Is AnyLanguageModel
Released on October 20 2026, AnyLanguageModel 1.0 is a drop-in replacement for Apple's Foundation Models framework. Developers change the import statement from FoundationModels to AnyLanguageModel and their existing session code continues to work.
The library supports local inference with llama.cpp, MLX and Core ML as well as multiple hosted providers. On iOS, macOS and visionOS 27 it also works with any model that conforms to Apple's LanguageModel protocol.
New Capabilities in Version 1.0
Tool calling and guided generation now function with every provider, including nested Generable types and arrays. Image support follows the Foundation Models syntax for attaching images to prompts, even with the system model.
Hosted support includes OpenAI Chat Completions and Responses, Anthropic Messages with adaptive thinking, Google's Gemini, Ollama Private Cloud Compute on OS 27, and Hugging Face Inference Providers via custom base URLs.
API keys are supplied through a closure so apps can fetch short-lived tokens from a server. The same code also runs on Linux. When switching providers, reasoning blocks are automatically omitted if unsupported.
Practical Local Model Examples
Popular small open-weight models include Google's Gemma 4 E2B in GGUF at roughly 3.4 GB quantized, Liquid AI's LFM2.5 2.6B at about 1.7 GB, and Alibaba's Qwen3.8 27B for MLX at 16 GB 4-bit. The first two fit on an iPhone; the Qwen model runs on Macs with 32 GB or more memory.
Selecting an MLX model takes one line of code. Larger models such as GLM-5.3 that require 1.5 TB can be accessed via hosted Inference Providers using the same session API. swift-huggingface and swift-xet speed up downloads from the Hub.
Why it matters for uncensored AI
Our analysis starts from the source fact that developers can now use the identical Swift code with local open-weight models via llama.cpp or MLX, or with hosted open-weight models through Inference Providers.
This gives Apple-platform developers a practical choice between Apple's system model and open-weight alternatives they select themselves. The library therefore expands the set of models that can be integrated into apps without changing core session logic.
For those seeking greater control over the underlying model in their own tools or companions, this reduces platform lock-in to a single vendor model. We recommend considering Muah AI for its multimodal approach to uncensored chat, photos, voice and phone calls when exploring hosted companion options that complement local development flexibility.
