AnyLanguageModel 1.0 Unifies Local and Hosted AI for Apple Apps
AnyLanguageModel 1.0 unifies local and hosted AI in Swift, adding tools, images, guided output, OS 27 support and faster Hugging Face model downloads.
Summary
Published October 20, 2026, AnyLanguageModel 1.0 arrives one year after its debut last November and dozens of releases, alongside swift-huggingface 1.0 and swift-xet 1.0. The drop-in Foundation Models replacement retains session code across Apple’s system model, open-weight models through llama.cpp, MLX and Core ML, and hosted services. Every provider supports tool calling and guided generation, including nested @Generable types and arrays. Image prompts use Foundation Models 27 syntax, including Apple’s system model on iOS, macOS and visionOS 27. OS 27 also accepts any LanguageModel implementation, including Core AI, while the companion packages accelerate Hugging Face Hub downloads.
Hosted support covers OpenAI Chat Completions and Responses, Open Responses endpoints, Anthropic Messages with adaptive thinking and effort, Google Gemini, Ollama, Private Cloud Compute on OS 27, Hugging Face Inference Providers, OpenRouter and custom OpenAI-compatible endpoints. Local and hosted backends share guided output, streaming tool calls, token usage and transcript reasoning; unsupported reasoning blocks are omitted when switching providers. API keys come from closures, enabling short-lived server tokens, and the same code runs on Linux. Hugging Face can host oversized models such as openai/gpt-oss-120b.
Google’s Gemma 4 E2B is about 3.4 GB at Q4_0 with quantization-aware training, while Liquid AI’s LFM2.5 2.6B is about 1.7 GB at Q4_K_M; both fit an iPhone. Alibaba’s 4-bit Qwen3.8 27B is about 16 GB and requires a Mac with at least 32 GB of memory. Z.ai’s GLM-5.3 needs about 1.5 TB in BF16 across four maxed-out Mac Studios, or hosted inference. Existing apps have migration guidance in all three release notes. Models that decide and listen are coming next.
Positives
- AnyLanguageModel 1.0 preserves Foundation Models session code across Apple system models, llama.cpp, MLX, Core ML and hosted providers.
- Tool calling and guided generation now work with every provider, including nested @Generable types and arrays.
- Foundation Models 27 image syntax supports prompts containing images on iOS, macOS and visionOS 27.
- swift-huggingface 1.0 and swift-xet 1.0 accelerate model downloads from the Hugging Face Hub into Swift apps.
- Gemma 4 E2B and LFM2.5 2.6B occupy about 3.4 GB and 1.7 GB respectively, allowing both to fit on iPhones.
- Linux support lets developers reuse AnyLanguageModel code in server deployments.
Risks & concerns
- Qwen3.8 27B requires about 16 GB of storage and a Mac with at least 32 GB of memory.
- GLM-5.3 requires roughly 1.5 TB in BF16 and four maxed-out Mac Studios unless accessed through hosted inference.
- Reasoning blocks are omitted from requests when a newly selected provider cannot replay them.
- Existing AnyLanguageModel, swift-huggingface and swift-xet apps may require code changes covered by separate migration guidance.
- Details remain limited for the forthcoming models designed to decide and listen.