The Pulse
Hugging Face Releases AnyLanguageModel 1.0 for Apple Apps
Hugging Face released AnyLanguageModel 1.0 on October 20, 2026, giving Swift developers a Foundation Models-compatible API for local and hosted language models. The package supports runtimes including MLX, llama.cpp and Core ML, alongside A
AI.info Team · Written with AI · fact-checked ·
“A shared API makes it easier to choose the right model for each task.”
Mattt
Hugging Face released AnyLanguageModel 1.0 on October 20, giving Apple-platform developers one Swift interface for models that run on a device and services that run remotely. The package is designed as a drop-in replacement for Apple’s Foundation Models framework: developers can change an import statement and keep the same session code while switching models or providers.
The release also brings swift-huggingface and swift-xet to version 1.0. Hugging Face says the two companion libraries support accelerated model downloads from its Hub to Swift apps. The company introduced AnyLanguageModel in November 2025 after developers described the effort of supporting different inference providers as a barrier to trying local models.
One import connects different model runtimes
AnyLanguageModel uses the Foundation Models API as its common interface. The same session code can work with Apple’s system model or open-weight models through llama.cpp, MLX and Core ML. On operating system version 27, the package also supports models that conform to Foundation Models’ LanguageModel protocol, including Core AI models.
The 1.0 release extends features across providers. Tool calling and guided generation—including nested Generable types and arrays—work across the supported providers, according to Hugging Face. Image prompts use Foundation Models 27 syntax, including with the system model on iOS, macOS and visionOS 27.
Cloud providers share the same session code
Developers can also use OpenAI’s Chat Completions and Responses APIs, Anthropic’s Messages API, Google’s Gemini API, Ollama, Open Responses and Hugging Face Inference Providers. Private Cloud Compute support is listed for operating system version 27. OpenAI-compatible services such as OpenRouter can connect through a custom base URL.
That option matters when a model is too large for the target device. Hugging Face’s example routes a request for OpenAI’s gpt-oss-120b through its Inference Providers endpoint while retaining the package’s session interface. Hosted providers also support guided generation, streamed tool calls and token usage reporting.
The package handles some differences between providers, too. Reasoning blocks are omitted from a request when the destination provider cannot replay them. Developers can supply API keys through a closure, allowing an app to request short-lived tokens from its own server rather than embedding a fixed key in the app.
Model size still determines the device
Hugging Face’s examples show the hardware trade-offs behind the unified interface. The post lists Gemma 4 E2B at about 3.4 GB in GGUF Q4_0 format and Liquid AI’s LFM2.5 2.6B at about 1.7 GB in GGUF Q4_K_M; the company says both fit on an iPhone. Alibaba’s Qwen3.8 27B in MLX format is about 16 GB at four bits and requires a Mac with at least 32 GB of memory, according to the post.
For a much larger model, Hugging Face cites Z.ai’s GLM-5.3 at about 1.5 TB in BF16 across four fully configured Mac Studios, while pointing developers to hosted inference for access. The figures make clear that a common API does not erase memory limits or the distinction between local and remote processing.
Version 1.0 sets the interface, not the hardware
For an existing Swift app, the published dependency example specifies version 1.0.0 and an optional MLX trait. Hugging Face directs developers upgrading from earlier releases to migration guidance for AnyLanguageModel, swift-huggingface and swift-xet. The post also says AnyLanguageModel runs on Linux, allowing the same code to be used on a server.
The release’s practical promise is a stable place in the app code to choose among Apple’s system model, local open-weight runtimes and hosted services. Which option an app can use still depends on model size, device memory, provider support and—when requests leave the device—how the developer manages credentials and routing.