News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
imgApple Foundation Models

Apple Foundation Models

Apple's 2025 on-device and server language models for Apple Intelligence

Apple Foundation Models is a family of generative language models built by Apple to power Apple Intelligence features across its software platforms. Introduced at the 2025 Worldwide Developers Conference, it pairs a compact on-device model of about 3 billion parameters with a mixture-of-experts server model designed for Private Cloud Compute.

Share on WhatsApp

  • On-device model has approximately 3 billion parameters
  • Server model uses a parallel track mixture-of-experts design
  • KV cache sharing cuts cache memory use by 37.5%
  • Vision encoders: ViT-g (1B) on server, ViTDet-L (300M) on device
  • Trained on Applebot crawls spanning hundreds of billions of pages

What it is

Apple Foundation Models is the set of generative language models behind Apple Intelligence, the AI features built into Apple's apps and operating systems. Apple introduced the current generation at its 2025 Worldwide Developers Conference. The family has two members: a compact model of roughly 3 billion parameters that runs on the device itself, and a larger mixture-of-experts model that runs on Apple's Private Cloud Compute servers. Both are tuned for Apple silicon and support 15 languages. They handle text and image input and are meant to power everyday features such as writing help, summarising and tool use, with an emphasis on keeping user data private.

How it came about

Apple presented the models at WWDC 2025 alongside a Foundation Models framework that lets app developers call the on-device model directly. On 17 July 2025 the company published a technical report and updated its research overview to match the model versions and evaluations in that paper. The report lists a large Apple author team led by researchers including Ethan Li, Anders Boesen Lindbo Larsen and Chen Zhang. A revised version of the paper appeared on 27 August 2025. Apple describes these models as part of a wider family of generative models built to support its users.

Inside the model

The on-device model is split into two blocks at a 5:3 depth ratio, with the second block reusing the key-value caches from the first, which Apple says cuts cache memory use by 37.5% and speeds up the first token. The server model uses a parallel track mixture-of-experts design: several smaller transformers, or tracks, run tokens independently and only sync at block boundaries, each block carrying its own expert layers. For long inputs, attention layers alternate between sliding-window local attention with rotary position embeddings and a global layer with no position embeddings. Vision comes from a 1B-parameter ViT-g encoder on the server and a 300M ViTDet-L with a Register-Window mechanism on device. Training data mixes licensed publisher content, open datasets and Applebot web crawls, filtered for personal data and unsafe material.

What it can do

Apple reports improved tool use and reasoning over earlier versions, plus image and text understanding, faster inference and support for 15 languages. The company says the models are optimised for efficiency on Apple silicon, with the on-device variant aimed at low-latency use and the server variant at harder tasks. Apple's overview compares the models against unnamed comparable models in its evaluations, and the figures were updated in July 2025 to match the technical report. The company also states that more languages will arrive later in the year.

How to use it

The models run inside Apple Intelligence on Apple devices and through Private Cloud Compute. At WWDC 2025 Apple released a Foundation Models framework giving developers direct access to the on-device model from their own apps. Apple has not released the weights; the models are proprietary and reached users through Apple's software rather than a public download.

Safety and testing

Apple says Responsible AI principles are applied throughout development. Training excludes users' private personal data and interactions unless a user explicitly opts in, and filters remove categories of personally identifiable information, profanity and unsafe material. Web crawling follows robots.txt so publishers can opt out while still appearing in Siri and Spotlight search, with fine-grained controls over which pages Applebot may use.

Reception and debate

Apple positions the models around privacy, noting that most processing happens on device or inside Private Cloud Compute rather than on external servers. The company's own account emphasises efficiency gains and expanded capabilities over the previous generation.

Why it matters

The Foundation Models framework opened the on-device model to third-party developers, extending Apple Intelligence beyond Apple's own apps. Apple frames the pair of models as a complementary suite covering both local, low-latency use and heavier server-side work, and says they belong to a larger family of generative models it has built for its platforms.

Questions people ask about Apple Foundation Models

What are Apple Foundation Models?

They are the generative language models Apple built to power Apple Intelligence. The family includes a roughly 3-billion-parameter model that runs on the device and a larger mixture-of-experts model that runs on Apple's Private Cloud Compute servers. Both were introduced at WWDC 2025.

Can I download or fine-tune the models?

No. Apple has not released the weights. Developers can reach the on-device model through the Foundation Models framework Apple introduced at WWDC 2025, and users get the models through Apple Intelligence features in Apple's software.

What data were they trained on?

Apple says training used licensed publisher content, curated open datasets and web pages crawled by its Applebot crawler, spanning hundreds of billions of pages. It also used image-text pairs, over 10 billion after filtering. Apple says it does not train on users' private data unless they opt in.

What makes the server model different?

It uses a parallel track mixture-of-experts architecture. Several smaller transformers process tokens independently and only synchronise at the input and output edges of each block, with expert layers inside each block. Apple says this lowers synchronisation overhead and keeps latency low while scaling.

Which languages does it support?

Apple says the 2025 models are designed to support 15 languages, with additional languages planned to become available later in the year.

Good reading and sources

    Where this page comes from

    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/apple-foundation-models-2025-update, arxiv.org/abs/2507.13575. 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.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.