Apple revealed groundbreaking AI models that will power Apple Intelligence features across iPhones, iPads, and Macs at WWDC 2025. These include both on-device and cloud-based models designed to work seamlessly with Apple’s ecosystem.
Now, Apple has released a detailed technical report titled “Apple Intelligence Foundation Language Models Tech Report 2025,” breaking down how these models function, how they were trained, and what kind of data they use.
What Are Apple’s New AI Models?
Apple introduced two core AI models to drive Apple Intelligence features in apps and services:
- A 3-billion-parameter model optimized for on-device performance on Apple Silicon.
- A more powerful cloud-based model running on Apple’s Private Cloud Compute (PCC) infrastructure.
Both models are multimodal (handling text and images) and multilingual, supporting multiple global languages. They were trained using a wide range of data sources including licensed content, web crawlers, and synthetically generated data.
How Do These AI Models Work?
On-Device Model
Apple split its on-device AI model into two blocks to reduce memory usage and improve speed:
- Block 1: Performs the majority of computations.
- Block 2: Skips some steps to save memory and boost speed.
This setup allows for quicker responses, ideal for real-time tasks like text prediction and summarisation.
Cloud-Based Model
Apple’s server model uses a high-performance architecture called Parallel-Track Mixture-of-Experts (PT-MoE). Here’s how it works:
- Instead of activating the whole model for every task, PT-MoE picks specific “expert” mini-models best suited for the job.
- For example, a travel-related query activates travel-trained experts only.
- This leads to faster, more efficient processing.
Apple also built a new Transformer architecture that processes various parts of a request in parallel to minimize delays and enhance performance.
What Are the Key Benefits of Apple’s AI Models?
One of the standout improvements is enhanced multilingual support. Apple increased the share of non-English training data from 8% to 30%, and expanded the vocabulary from 100,000 to 150,000 tokens.
This enables Apple Intelligence to understand and respond more accurately in various languages. Apple even tested performance with native speakers, ensuring cultural and linguistic fluency.
Other benefits include:
- Improved performance of features like Writing Tools in more languages.
- Third-party developers can now use Apple’s on-device model to add AI summarisation and rewriting features directly into their apps, without user data ever leaving the device.
Where Did Apple Get Its Training Data?
Apple claims it trained its AI models on high-quality, diverse datasets, without using any personal device data. Instead, Apple relied on:
- Publisher-licensed content
- Public and open-source information
- Web data via Applebot, its proprietary web crawler
Text Data Collection
Applebot crawled billions of web pages, including dynamic and language-diverse sites. The content was filtered using AI-based quality control, which helped retain valuable information while avoiding low-quality or offensive material.
Image Data Collection
To help models understand images, Apple used:
- Licensed and public photos with captions
- AI-generated image captions
- Infographics, tables, and charts
Apple even generated synthetic training examples such as creating a chart and then generating questions based on it to enrich training.
Final Thoughts
Apple’s AI models represent a significant leap in on-device and cloud-based intelligence. With strong privacy practices, global language support, and developer accessibility, these models are shaping the future of intelligent computing on Apple platforms.