庞若鸣交班陈智峰，苹果发布2025基础模型技术报告

Core Viewpoint - Apple has released a technical report on its Apple Intelligence foundational language models for 2025, showcasing advancements in model architecture, training methods, and performance evaluations compared to similar models [2][4]. Model Innovations - Apple introduced two foundational language models: a 3 billion parameter device model optimized for Apple chips and a scalable cloud model utilizing a new parallel track mixture of experts (PT-MoE) Transformer architecture [6][11]. - The PT Transformer architecture allows for parallel execution of smaller Transformer modules, reducing synchronization overhead and improving training and inference latency [8][12]. Visual Understanding - A visual encoder has been integrated to extract visual features from input images, enhancing the model's ability to understand images and perform tool calls [9][10]. - The device model employs a 300 million parameter visual backbone, while the server model consists of 1 billion parameters, both designed to capture fine-grained local details and global context [10]. Developer Framework - Apple has launched a new Swift-based foundational model framework that includes guided generation, constrained tool calls, and LoRA adapter fine-tuning, enabling developers to easily integrate these features [21][22]. - The framework supports a device-side language model with approximately 3 billion parameters, excelling in various text tasks such as summarization and entity extraction [22]. Responsible AI Practices - Apple emphasizes its commitment to responsible AI, implementing content filtering and regional customization assessments to ensure user privacy and safety [23]. Leadership Transition - Following the release of the report, Ruoming Pang expressed gratitude to contributors and passed the leadership baton to Zhifeng Chen and Mengyu Li, indicating a shift in the management structure of Apple's AI team [24][26].