size of foundational models
Asked on 06/11/2025
1 search
The foundational models discussed at Apple's WWDC are designed to run on devices and have around 3 billion parameters. This is significantly smaller than server-based large language models (LLMs) like ChatGPT, which have hundreds of billions of parameters. The on-device model is optimized for tasks such as summarization, extraction, and classification, but it is not designed for tasks requiring extensive world knowledge or advanced reasoning. This smaller size allows the model to run efficiently on devices like iPhones and iPads, maintaining privacy and offline capabilities.
For more details, you can refer to the session Explore prompt design & safety for on-device foundation models (03:01) and Meet the Foundation Models framework (02:57).
Explore prompt design & safety for on-device foundation models
Design generative AI experiences that leverage the strengths of the Foundation Models framework. We’ll start by showing how to design prompts for the on-device large language model at the core of Apple Intelligence. Then, we’ll introduce key ideas around AI safety, and offer concrete strategies to make your generative AI features safe, reliable, and delightful.
Meet the Foundation Models framework
Learn how to tap into the on-device large language model behind Apple Intelligence! This high-level overview covers everything from guided generation for generating Swift data structures and streaming for responsive experiences, to tool calling for integrating data sources and sessions for context management. This session has no prerequisites.
Deep dive into the Foundation Models framework
Level up with the Foundation Models framework. Learn how guided generation works under the hood, and use guides, regexes, and generation schemas to get custom structured responses. We’ll show you how to use tool calling to let the model autonomously access external information and perform actions, for a personalized experience. To get the most out of this video, we recommend first watching “Meet the Foundation Models framework”.