Apple has unveiled its 2025 Apple Intelligence foundational language model technology report, showcasing two groundbreaking multilingual, multimodal foundational models. One is a 3B-parameter device model tailored for Apple chips, and the other is a scalable cloud model. The cloud model leverages an innovative PT-MoE Transformer architecture integrated with a global-local attention mechanism to enhance inference efficiency. The device-side model, meanwhile, employs a ViTDet-L vision backbone and a registration window mechanism to bolster visual understanding capabilities. Furthermore, Apple has introduced the Swift Core Model Framework, which integrates guided generation, tool invocation, and LoRA fine-tuning functions, empowering developers to swiftly create AI applications. These new models excel in multitask and multilingual processing, balancing low-latency performance on devices with high-performance requirements in the cloud, while adhering to stringent privacy protection measures and responsible AI principles. Notably, Apple's AI team is now under the leadership of Chen Zhifeng, following the departure of previous head Pang Ruoming to join Meta.