Apple released a tech report detailing its new on-device and cloud-based foundation models. The on-device model, with 3 billion parameters, is split into two blocks for improved memory efficiency and speed. The cloud-based model uses a custom Parallel-Track Mixture-of-Experts architecture for faster, more accurate responses. Multilingual support significantly improved, increasing multilingual data from 8% to 30% of training data. Data sources included web crawling (respecting robots.txt), licensed content, and synthetic data. The report highlights Apple's focus on privacy and efficiency.

Prepared by Jonathan Pierce and reviewed by editorial team.