Apple has revealed key details about its latest AI models powering the Apple Intelligence features on iPhones, iPads, and Macs. In a new technical report titled Apple Intelligence Foundation Language Models Tech Report 2025, the company explained how the models function and where their training data comes from.
Apple’s AI system consists of two foundation models: a lightweight 3-billion-parameter model running directly on Apple Silicon devices, and a more powerful server-based model hosted on Apple’s Private Cloud Compute (PCC). Both models are multilingual and multimodal, capable of understanding multiple languages and handling both text and image inputs.
The on-device model is split into two components to minimize memory usage and speed up response time—ideal for real-time tasks like summarisation and suggestions. The cloud-based model uses a Parallel-Track Mixture-of-Experts (PT-MoE) architecture, which intelligently routes tasks to specialized “experts” trained on different content types, enhancing both speed and accuracy.
A major highlight of the new models is expanded multilingual support. Apple increased non-English language data from 8% to 30% and expanded the vocabulary from 100,000 to 150,000 tokens. This enables Apple Intelligence to perform better in multiple languages, verified through native speaker testing.
Apple has also opened its on-device model to third-party developers, allowing them to use AI features within their apps without sending user data off the device.
As for data sources, Apple stated that it does not use personal data or private device activity. The models were trained on licensed content, public/open-source data, and web pages crawled by Applebot. The crawler follows “robots.txt” instructions, giving publishers control over their inclusion.
Apple also used a combination of licensed and public images, along with synthetic image-text data, to improve the AI’s visual understanding.