Mirror Review
July 22, 2025
Summary:
- Apple has developed two main AI models, known as foundation language models, to run its new “Apple Intelligence” features.
- One model is small enough to run directly on your device for speed and privacy, while a second, more powerful model runs on Apple’s secure cloud servers for complex tasks.
- Apple emphasizes its commitment to privacy by stating that it does not use its users’ private data or interactions to train these models.
How do you make artificial intelligence both personal and private?
Apple believes it has found the answer!
At its 2025 Worldwide Developers Conference, Apple unveiled its new Apple Foundation language models.
These powerful models are the core engine behind “Apple Intelligence,” a new set of features designed to integrate generative AI directly into the apps and experiences people use every day, all while protecting their privacy.
The Two Main Language Models:
Apple has developed two distinct, but complementary models to handle different kinds of tasks.
1. The On-Device Model:
This is a compact, roughly 3-billion-parameter model optimized to run efficiently on Apple silicon. It’s designed for speed and low latency, handling many requests directly on your iPhone or Mac without needing to send data to the cloud.
2. The Server Model:
For more demanding tasks that require greater power, Apple uses a larger, scalable server model. This model runs on Apple’s “Private Cloud Compute“ platform, ensuring that even data processed in the cloud remains secure and private.
What Can These Apple Language Models Do?
These models are built to be genuinely helpful across a wide range of activities. They are multimodal, meaning they can understand both text and images, and have been designed to support 15 languages.
Their capabilities include:
- Text summarization and understanding
- Generating creative content
- Image understanding and reasoning
- Executing “tool calls” to perform actions in apps
The on-device model excels at tasks like refining your writing, extracting key information, and engaging in short dialogues, while the server model handles more complex reasoning.
How Were These Apple Language Models Trained
The quality of an AI model is heavily dependent on the data it’s trained on. Apple used a mix of sources for its training data:
- Licensed information from publishers
- Publicly available data crawled by its web-crawler, Applebot
- High-quality synthetic data created in-house
Crucially, Apple makes it clear that it does not use its users’ private personal data or their interactions when training its foundation models.
They also respect publisher requests to opt out of having their content used for training.
To handle visual tasks, the models were trained on a massive dataset of over 10 billion image-text pairs and 175 million documents that contained interleaved images and text.
Key Innovations in the Apple Foundation Language Models
To achieve high performance and efficiency, Apple developed several new technologies:
- On-Device Efficiency:
The on-device model uses a technique called KV-cache sharing, which cleverly reduces memory usage by 37.5% and makes the model respond faster.
- Server Scalability:
The server model features a novel “Parallel-Track Mixture-of-Experts” (PT-MoE) architecture. This design allows different parts of the model to work independently, significantly reducing processing delays and improving efficiency without sacrificing quality.
- Model Compression:
To make the models even smaller and faster, Apple uses advanced compression. The on-device model’s weights are compressed to just 2 bits, while the server model uses a method called ASTC, originally developed for graphics, which is decoded by dedicated hardware on Apple GPUs.
A New Toolkit for Developers
Apple is also empowering app developers to build their own AI features with the new “Foundation Models framework.” This framework gives developers direct access to the ~3B-parameter on-device model.
A highlight is “guided generation,” an intuitive way for developers to get structured output from the model using the Swift programming language.
The framework also supports “tool calling,” allowing developers to connect the model to their app’s specific functions and data sources. For advanced use cases, a Python toolkit is available to train custom skills for the model.
Evaluation and Responsible AI
Apple conducted extensive evaluations and built its models with safety and its core values in mind.
- Performance:
In human evaluations, Apple’s on-device model performed favorably against similarly sized models. The server model was found to be competitive but behind larger proprietary models like GPT-4o.
- Responsible AI:
Apple’s approach is guided by four principles: Empower users, Represent users, Design with care, and Protect privacy. The company has built-in safety guardrails to filter harmful content and has developed evaluation datasets to specifically test for risks and biases across different languages and cultures.
Looking Ahead
The Apple Intelligence Foundation Language Models launch moves beyond a single chatbot to integrate AI deeply and thoughtfully across its entire ecosystem.
The focus on on-device processing and Private Cloud Compute establishes privacy as a key differentiator in the competitive AI landscape.
By opening up the Apple Foundation language models to developers, Apple is channeling a new wave of intelligent, private, and helpful app experiences that could fundamentally change how we use our devices every day.
