Apple Foundation Models (2025)

What it is

Apple Foundation Models (2025) are the second generation of the language models that Apple uses for Apple Intelligence. Apple announced them on 9 June 2025 at WWDC, its annual developer conference, and published a technical report on 17 July 2025 ^[1]. As with the first generation, there are two main models: an on-device model that runs directly on Apple hardware and a server model that runs in Apple's cloud infrastructure.

A language model is software trained on large amounts of text to read and produce language. Apple's models in this generation can also read images, which is the main new capability.

Origins

The first generation, the models behind Apple Intelligence at its launch, was the starting point. The 2025 models were a revision of that design rather than a replacement from scratch. Apple's own report described the changes made to the architecture, the training data and the way the models run on devices ^[1].

Key features

The on-device model has about 3 billion parameters, a count of the numbers inside the network that are adjusted during training. Its architecture is split into two blocks at a 5:3 depth ratio. The second block reuses the key-value (KV) cache of the first block. The KV cache holds intermediate results so the model does not have to recompute them for every new word, and sharing it cuts the memory needed for the cache by 37.5% ^[1]. This matters on a phone, where memory is limited.

The on-device model is compressed to 2 bits per weight using quantization-aware training, a method that trains the model with the compression in mind so that the loss in quality is smaller ^[1].

The server model uses a new design, Parallel-Track Mixture-of-Experts (PT-MoE). In a mixture-of-experts model, each input is sent to a small set of specialist sub-networks, called experts, rather than through the whole network. In PT-MoE, several smaller transformer tracks process tokens independently and only synchronise at their boundaries, and each track has its own MoE layers ^[1]. A transformer is the neural network design that most current language models use.

For long documents, Apple continued pre-training at up to 65K tokens, with sliding-window attention and global attention interleaved, so the model can handle long input ^[1].

Both models support 15 languages and gained image understanding. The server model uses a 1-billion-parameter ViT-g image encoder, and the on-device model uses a 300-million-parameter ViTDet-L encoder ^[1]. An image encoder turns a picture into numbers the language model can work with.

Training data included Applebot web data, licensed publisher data, and more than 10 billion image-text pairs plus more than 5 billion synthetic image captions ^[1]. Synthetic captions are descriptions generated by a model rather than written by people.

Apple's human evaluations compared the models with other systems. The on-device model compared favourably with Qwen-2.5-3B and was competitive with Qwen-3-4B and Gemma-3-4B in English. The server model compared favourably with Llama 4 Scout but trailed Qwen-3-235B and GPT-4o ^[1].

Developer access

The Foundation Models framework lets developers use the on-device model from Swift, Apple's programming language for its platforms. It supports guided generation, marked with the @Generable attribute, which asks the model to return output in a fixed structure, and tool calling, which lets the model use functions the app provides. Inference is free of charge ^[1]. The framework shipped with iOS 26 on 15 September 2025 ^[2].

Where it sits

This generation is the link between the first Apple models and the next one. Apple's third generation, AFM 3, was published on 8 June 2026 and is a family of five models rather than two ^[3].