Apple opens its on-device model to developers with second-generation foundation models

Background

Apple builds its own foundation models, the large language models that sit behind Apple Intelligence. The first generation was the set of models Apple used when Apple Intelligence arrived. The second generation was the next version of that work, and it came with a change of access for outside developers. The occasion was WWDC, the Worldwide Developers Conference, where Apple presents its software plans for the coming year.

What happened

On 9 June 2025, at WWDC, Apple introduced its second-generation foundation models and opened the on-device model to third-party apps ^[1][2]. Developers who write apps for Apple platforms could, for the first time, build features on top of the same model that Apple uses inside its own system apps.

Apple published the accompanying technical report on 17 July 2025, more than a month after the announcement ^[1]. The report gave the design and training details of the models.

Key details

The on-device model has about 3 billion parameters, and the server model uses a Parallel-Track Mixture-of-Experts architecture ^[1]. A parameter is one of the numbers inside a neural network that are set during training. A mixture-of-experts design sends each input to a few specialist parts of the network, which can keep the cost of running a large model lower than running every part for every input.

Both models gained image understanding and support for 15 languages ^[1]. The new image understanding lets the model answer questions about pictures as well as text, which was not part of the first generation's core abilities.

The Foundation Models framework lets Swift developers call the model with two features. Guided generation asks the model to produce output in a set structure, so an app can rely on the form of the answer. Tool calling lets the model use functions that the app defines ^[1]. Inference is free of charge for developers ^[1], so an app that uses the model does not add a per-request cost for the model itself.

The framework and Live Translation shipped in iOS 26 on 15 September 2025 ^[2].

Impact

The main change for developers was access. The on-device model had been part of Apple's own system, and the framework made it available to third-party apps. A developer who wants to add an AI feature to an app can now call the model from Swift without building or hosting a model of their own.

For users, the change means that the apps they already have can offer AI features tied to the operating system. For readers who want to learn the field, the event also shows a pattern that is common across the industry: a company keeps a large model for its own products and opens a model to outside builders. The technical report of July gives the full design and training detail for anyone who wants to read beyond the announcement.