TL;DR
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Hugging Face has released AnyLanguageModel 1.0, a Swift library that lets developers use a shared Foundation Models-style API with local and hosted language models. The release also includes 1.0 versions of swift-huggingface and swift-xet; Hugging Face says upcoming work will add models that decide and listen.
Hugging Face has released AnyLanguageModel 1.0, a Swift library that gives app developers a shared API for using on-device and hosted language models in apps for Apple platforms. The release arrives about a year after the project was introduced and is accompanied by 1.0 versions of swift-huggingface and swift-xet, tools for downloading models from the Hub in Swift apps.
AnyLanguageModel is designed as a drop-in replacement for Apple’s Foundation Models framework interface. Developers can change the import statement and keep the same session code while selecting among Apple’s system model, open-weight models running through llama.cpp, MLX or Core ML, and hosted services. Hugging Face says the library now supports tool calling and guided generation across providers, including nested generated types and arrays.
The update adds image prompts using Foundation Models 27 syntax, including use of the system model on iOS, macOS and visionOS 27. On OS 27, the library can also work with models that conform to Apple’s LanguageModel protocol, including Core AI models. Hugging Face says hosted models support guided generation, streaming tool calls and token-usage reporting, and that reasoning blocks are included in transcripts but omitted from requests when a newly selected provider cannot replay them.
Hosted integrations include OpenAI’s Chat Completions and Responses APIs, Open Responses-compatible endpoints, Anthropic’s Messages API, Google’s Gemini API, Ollama, Private Cloud Compute on OS 27, and Hugging Face Inference Providers. The library also runs on Linux. Hugging Face has changed API-key handling so an app can retrieve short-lived credentials through a closure, including from a developer’s own server.
Hugging Face · One-year release · Apple platforms
One Swift API.
More ways to run AI.
AnyLanguageModel 1.0 brings local and hosted language models behind a shared Foundation Models-style interface, one year after the project began.
“Local and hosted inference each have their strengths, and both have a role in building this new generation of intelligent apps.”
— Hugging Face01 / What shipped
A shared interface, a wider model menu
Teams can keep a familiar session pattern while choosing the model and provider that fit a task. Hugging Face describes the library as a drop-in alternative to Apple’s Foundation Models interface.
Choose where it runs
Use Apple’s system model, open-weight models through llama.cpp, MLX or Core ML, and hosted services within one app architecture.
Richer interactions
Tool calling and guided generation work across providers, including nested generated types and arrays. Image prompts use Foundation Models 27 syntax.
Three releases at 1.0
AnyLanguageModel, swift-huggingface and swift-xet reached 1.0. The download libraries help Swift apps retrieve models from the Hub.
Retrieve
Download models from the Hugging Face Hub with Swift tools.
Select
Pick a system, local open-weight or hosted model.
Interact
Use a shared session API for generation and tool calls.
Adapt
Change providers as task, device and deployment needs change.
02 / Model scale
From phone-sized to provider-hosted
These are Hugging Face’s example model formats and sizes. They do not guarantee comparable speed, memory use or output quality across devices.
Gemma 4 E2B
Q4_0 GGUF · cited as fitting on an iPhone
LFM2.5 2.6B
Q4_K_M GGUF · cited as fitting on an iPhone
Qwen3.8 27B
Four-bit MLX · Macs with at least 32 GB memory
GLM-5.3
BF16 example described as available through Inference Providers
Model size changes the deployment choice
Approximate scale shown on a logarithmic visual: the largest example is hundreds of times bigger than the phone-sized examples.
Actual requirements depend on format, quantization, device and configuration. Local and hosted choices also differ in latency, cost and data handling.
03 / Provider reach
One API, many integrations
Hosted connections cover several API families, while local backends and Apple system models offer on-device paths. Linux is also supported.
Multiple services
OpenAI Chat Completions and Responses, Open Responses-compatible endpoints, Anthropic Messages, Google Gemini and Ollama.
Inference Providers
Access hosted models through Hugging Face. The update also includes streaming tool calls and token-usage reporting.
System and Core AI
Image prompts include system-model support on iOS, macOS and visionOS 27. OS 27 adds models conforming to Apple’s LanguageModel protocol.
04 / What the announcement does not show
Useful infrastructure, with open questions
A shared API may reduce provider-specific code, but the release announcement does not establish production outcomes or quantify the benefit.
Benchmarks and adoption
No independent local-versus-hosted benchmarks, app performance measurements or developer adoption figures were provided.
Device, cost and quality
Memory use, response speed, output quality and hosted costs vary by model, provider, device and configuration.
Check provider details
Teams should review individual provider terms, OS requirements, release notes and migration guidance before deployment.
Short-lived API keys
Apps can retrieve credentials through a closure, including from a developer’s own server. Security practices still depend on implementation.
05 / Assessment
Promising foundation; production proof is next
Hugging Face says models that “decide” and “listen” are coming, without dates, feature details or device guidance.
A practical infrastructure release
Why it matters: Swift teams may have less provider-specific integration code to maintain and more room to match models to tasks. What remains unproven: a common API cannot erase differences in providers, latency, cost, device capability or model quality. Independent benchmarks and developer reports would clarify how well the abstraction holds up in deployed apps.
Models that decide
Planned direction from Hugging Face; no release date or feature description announced.
Models that listen
Also described as coming soon. Supported devices and practical capabilities remain unspecified.
One Swift Interface, More Model Choices
The release addresses a practical obstacle for teams building model features into Swift apps: supporting several inference providers can require separate integrations and code paths. A shared session interface can make it easier to change providers or use different models for different tasks, though the announcement does not quantify development time saved or report independent adoption figures.
Local inference can keep model processing on a user’s device, while hosted services can provide access to models too large to run locally. Hugging Face presents the library as a way to use both approaches within the same app architecture. That flexibility matters as model choice increasingly depends on a task’s resource needs and where inference should run.
The accompanying download libraries extend the effort beyond inference. Hugging Face says swift-huggingface and swift-xet provide accelerated downloads from the Hub to Swift apps. Together, the releases offer developers a connected path to retrieve models and use them through a common interface, although performance, device compatibility and operational costs will vary by model and provider.
As an affiliate, we earn on qualifying purchases.
A Year of Swift Model Support
Hugging Face introduced AnyLanguageModel the previous November, describing it as a drop-in alternative to Apple’s Foundation Models framework that lets developers select a model and provider. The project grew from feedback from app developers who were interested in local models but wanted a simpler way to support multiple inference providers.
Since that launch, the library has added local backends, hosted services and features such as tool calling and guided generation. The announcement points to a range of open-weight models as possible options: it lists Gemma 4 E2B at about 3.4 GB in Q4_0 GGUF and LFM2.5 2.6B at about 1.7 GB in Q4_K_M GGUF, which it says fit on an iPhone. It lists Qwen3.8 27B at about 16 GB in four-bit MLX form for Macs with at least 32 GB of memory.
Those sizes and device requirements are Hugging Face’s stated examples, not a guarantee of equivalent speed or quality across devices. The announcement also gives GLM-5.3 as an example of a model too large for ordinary local use: its BF16 version takes about 1.5 TB, according to Hugging Face, and is available through Inference Providers.
““Local and hosted inference each have their strengths, and both have a role in building this new generation of intelligent apps.””
— Hugging Face
As an affiliate, we earn on qualifying purchases.
Adoption and Device Results Are Unreported
The announcement does not provide independent benchmarks comparing local and hosted inference, measurements of app performance, or figures for developer adoption. The listed model sizes describe particular formats and quantization levels; actual memory use, response speed and output quality may differ by device and configuration.
It is also not clear from the announcement which providers or features have limitations beyond the stated OS requirements, or how costs compare across hosted services. Developers will need to check the individual provider terms and release notes for implementation details and migration guidance.
As an affiliate, we earn on qualifying purchases.
More On-Device Model Capabilities
Hugging Face says it plans to extend the same approach to other kinds of models, with “models that decide” and “models that listen” coming soon. The announcement does not give release dates, feature descriptions or supported devices for those capabilities.
For developers adopting version 1.0, the immediate next step is to review the migration guidance in the AnyLanguageModel, swift-huggingface and swift-xet release notes. The project’s future direction and the practical value of the new releases will become clearer as those capabilities arrive and developers report how the libraries perform in deployed apps.
Source: Hugging Face
As an affiliate, we earn on qualifying purchases.
Where I land
My assessment is that AnyLanguageModel 1.0 is a useful infrastructure release: it lowers the amount of provider-specific code a Swift team may need to maintain and makes local and hosted models easier to consider within one app. That could give developers more room to match a model to a feature without rebuilding their integration each time.
The strongest counterargument is that a common API cannot remove the real differences between providers, device capabilities, latency, cost and model quality. Hugging Face’s announcement describes features and example model sizes, but it does not establish that the library improves app performance or that switching providers is seamless in production.
I would raise my assessment if independent benchmarks and developer reports showed reliable behavior across a broad range of devices and providers, along with clear migration and security practices for hosted credentials. Evidence of compatibility problems, substantial integration work, or poor local performance on supported devices would lower it.
Key Questions
What is AnyLanguageModel 1.0?
It is a Swift library that provides a Foundation Models-style interface for using system, local open-weight and hosted language models in apps.
Which local model backends does it support?
Hugging Face lists llama.cpp, MLX and Core ML for open-weight models, as well as Apple’s system model. Support for models conforming to Apple’s LanguageModel protocol requires OS 27, according to the announcement.
Can developers use hosted models with the same code?
Hugging Face says developers can retain the same session code while selecting hosted providers, including OpenAI, Anthropic, Google Gemini, Ollama and compatible endpoints. Available features depend on the provider and its API.
What else was released with AnyLanguageModel 1.0?
Hugging Face released swift-huggingface 1.0 and swift-xet 1.0, which it describes as tools for accelerated model downloads from the Hub in Swift apps.
Source: Hugging Face
Halloween Picks
halloween
As an affiliate, we earn on qualifying purchases.
