Hugging Face AnyLanguageModel 1.0: one Swift API for local and hosted models

Hugging Face announced AnyLanguageModel 1.0 on its blog on 20 October 2026, along with 1.0 releases of swift-huggingface and swift-xet, which speed up model downloads from the Hub in Swift apps. AnyLanguageModel is an open-source Swift library that mirrors Apple's Foundation Models API, so an app can change which model answers without rewriting its session code. Hugging Face says the migration starts by replacing `import FoundationModels` with `import AnyLanguageModel`.

What it covers

According to the post, the same code can drive the Apple system model, open-weight models run through llama.cpp, MLX and Core ML, and hosted services. The hosted list is OpenAI's Chat Completions and Responses APIs, Open Responses endpoints, Anthropic's Messages API (with adaptive thinking and effort options), Google's Gemini API, Ollama, Private Cloud Compute on OS 27, and Hugging Face Inference Providers or any OpenAI-compatible endpoint such as OpenRouter through a custom base URL.

• Tool calling and guided generation (including nested @Generable types) work with every provider, and hosted providers get streaming tool calls and token usage on responses.
• Reasoning blocks are kept in the transcript and are dropped automatically from requests when a new provider cannot replay them, which matters if you switch models mid-conversation.
• API keys come from a closure rather than a string, so an app can fetch short-lived tokens from its own server; the README discusses when to let users bring their own keys and when to route through a proxy.
• The library also runs on Linux, so the same code can work on a server.

Which local models fit

Hugging Face names three of the most downloaded models on the Hub: Gemma 4 E2B from Google (about 3.4 GB in GGUF at Q4_0), LFM2.5 2.6B from Liquid AI (about 1.7 GB at Q4_K_M) and Qwen3.8 27B for MLX from Alibaba (about 16 GB at 4 bits). It says the first two fit on an iPhone and the third runs on a Mac with 32 GB of memory or more. Larger models, such as Z.ai's GLM-5.3 at about 1.5 TB in BF16, are reached through Inference Providers instead.

What the post leaves out

This is Hugging Face describing its own library. The post offers no latency, battery drain, answer-quality or cost comparison between local and hosted models, and no data on how many apps use the library. Image input depends on OS 27, and the Private Cloud Compute option is likewise limited to it. Hugging Face says models that decide and models that listen are coming next, without dates.

What to do

• If you ship an Apple-platform app with AI features, prototype one task against a small local model and one hosted model through the same session code, and compare quality, latency and cost per task on your own devices.
• Decide where API keys live before launch: use short-lived tokens from your server unless users must bring their own keys.
• Check the licence and data terms of each model and provider you route to; one shared API does not make their terms the same.