Apple has unveiled a significant expansion of its Artificial Intelligence strategy, announcing that its on-device Apple Intelligence large language model (LLM) will now be accessible to developers via the Foundation Models framework in upcoming releases of macOS ´Tahoe´ 26, iOS 26, and iPadOS 26. This move aims to empower developers to harness the capabilities of the Apple Intelligence system, facilitating guided content generation and tool invocation entirely on users´ devices, thus preserving privacy by avoiding the transfer of sensitive data to remote servers.
At the same time, Apple has upgraded its Xcode development platform to directly integrate OpenAI´s ChatGPT, enabling software engineers to leverage cloud-based LLMs for code generation, documentation, and other productivity enhancements. Notably, while Apple´s own LLM runs entirely on-device, ChatGPT and other third-party hosted models will require user data to be transmitted to external servers, emphasizing a clear distinction in how the company is approaching privacy versus functionality.
This announcement comes against the backdrop of a revealing paper published by Apple´s own researchers, who cast doubt on the real capabilities of so-called ´reasoning models.´ The study finds that large reasoning models (LRMs) exhibit inconsistent reasoning, fail at explicit algorithmic tasks, and lack generalizable problem-solving skills beyond moderate complexity. Both standard and reasoning models collapse when facing problems of higher complexity, challenging current marketing narratives and existing assumptions about generalizable Artificial Intelligence reasoning abilities in commercial LLMs. Despite these reservations, Apple is moving forward with beta releases of its new features through the Apple Developer Program, planning a public beta next month and a general release in the fall. The company and its developer ecosystem now face the task of balancing the promise of richer, more intuitive apps with a deeper awareness of the current limitations of reasoning in LLM-driven technologies.
