Apple Foundation Models
Apple's on-device and server language models behind Apple Intelligence
Apple Foundation Models are the language models Apple built for Apple Intelligence, its personal intelligence system in iOS 18, iPadOS 18 and macOS Sequoia. Unveiled at the 2024 Worldwide Developers Conference, the family pairs a roughly 3 billion parameter model running on the device with a larger one on Apple silicon servers, plus coding and diffusion models.
- On-device model of about 3 billion parameters, plus a larger server model
- Runs on iPhone 15 Pro at roughly 0.6 ms to first token and 30 tokens per second
- LoRA adapters let one base model specialise per task without retraining it
- Compressed to an average 3.7 bits per weight with no reported accuracy loss
- Trained on Apple's open-source AXLearn framework
What it is
Apple Foundation Models are the generative models Apple built into Apple Intelligence, its personal intelligence system woven into iOS 18, iPadOS 18 and macOS Sequoia. Apple introduced them at its 2024 Worldwide Developers Conference. The set includes a language model of roughly 3 billion parameters that runs on the device itself, and a bigger language model served from Apple silicon servers through Private Cloud Compute. The wider family also covers a coding model for Xcode and a diffusion model for making images in apps such as Messages. Apple fine-tunes them for everyday jobs: drafting and polishing text, sorting and summarising notifications, making images for chats, and carrying out actions inside apps.
How it came about
Apple showed the models at the 2024 Worldwide Developers Conference as the engine behind Apple Intelligence. On 29 July 2024 the company updated its published figures to match the model versions and evaluations in a technical report released the same day, titled Apple Intelligence Foundation Language Models. That paper lists a large author team led by Tom Gunter and Zirui Wang. Apple says more detail on the broader set of models, including the coding and diffusion ones, would follow.
Inside the model
Both language models use grouped-query attention and share one embedding table between input and output to save memory. The on-device model has a 49K vocabulary; the server model uses 100K, adding language and technical tokens. Training ran on Apple's open-source AXLearn framework, built on JAX and XLA, using data, tensor and sequence parallelism plus fully sharded data parallel training. Training data combined licensed material and web pages gathered by AppleBot, with filters for personal identifiers and low-quality text. Post-training used two new methods: rejection sampling fine-tuning with a teacher committee, and reinforcement learning from human feedback with mirror descent policy optimisation. Task-specific LoRA adapters sit on top of the frozen base model.
What it can do
Apple reports that adapters let one base model specialise on the fly for each feature while keeping its general knowledge intact. On an iPhone 15 Pro the on-device model reaches about 0.6 millisecond time-to-first-token per prompt token and generates 30 tokens per second, before any token speculation. Compression to an average 3.7 bits per weight, or more aggressively 3.5, is said to match the accuracy of the uncompressed model. Apple says its two post-training algorithms brought sizeable gains in instruction following.
How to use it
The models ship inside Apple Intelligence on iOS 18, iPadOS 18 and macOS Sequoia, with the larger model reached through Private Cloud Compute on Apple silicon servers. Apple has not released the weights. The AXLearn training framework is open source, and web publishers can opt out of having their pages used for Apple Intelligence training.
Safety and testing
Apple sets out Responsible AI principles covering user empowerment, representation, careful design and privacy. It says it looks for misuse and harm at every stage, from design through training to quality evaluation, and improves tools with user feedback. Personal data and user interactions are excluded from training unless a user explicitly opts in, and publicly posted personal identifiers are filtered out.
Questions people ask about Apple Foundation Models
What are Apple Foundation Models?
They are the generative models Apple built for Apple Intelligence. The set includes a roughly 3 billion parameter language model that runs on the device, a larger language model served from Apple silicon servers, a coding model for Xcode and a diffusion model for images.
Where do they run?
The smaller language model runs on the device itself, inside iOS 18, iPadOS 18 and macOS Sequoia. The larger one runs on Apple silicon servers and is reached through Private Cloud Compute, Apple's private cloud setup.
Can I download the weights?
No. Apple has not published the model weights. The AXLearn framework used to train them is open source, but the models themselves are delivered through Apple's own software.
What data were they trained on?
Apple says it used licensed data and publicly available web pages collected by its AppleBot crawler, with filters removing personal identifiers and low-quality text. User data and interactions are left out unless a person explicitly opts in, and publishers can opt out.
How does one model handle so many tasks?
Apple attaches small adapter modules, trained with LoRA, to layers of a shared base model. Only the adapters change per task, so the base weights stay fixed and the model can swap in the right adapter for the job at hand.
Good reading and sources
There is no English Wikipedia article on Apple Foundation Models yet, so the facts on this page come from the sources listed above: machinelearning.apple.com/research/introducing-apple-foundation-models, arxiv.org/abs/2407.21075. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.
It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.