苹果公司：2025 Apple Intelligence基础语言模型技术报告（英文版）（27页）.pdf

1、Apple Intelligence FoundationLanguage ModelsTech Report 2025AppleWe introduce two multilingual,multimodal foundation language models that powerApple Intelligence features across Apple devices and services:(i)a3B-parameteron-device model optimized for Apple silicon through architectural innovations s

2、uchas KV-cache sharing and 2-bit quantization-aware training;and(ii)a scalable servermodel built on a novel Parallel-Track Mixture-of-Experts(PT-MoE)transformer thatcombines track parallelism,mixture-of-experts sparse computation,and interleavedgloballocal attention to deliver high quality with comp

3、etitive cost on Apples PrivateCloud Compute platform.Both models are trained on large-scale multilingual andmultimodal datasets sourced via responsible web crawling,licensed corpora,andhigh-quality synthetic data,then further refined with supervised fine-tuning and rein-forcement learning on a new a

4、synchronous platform.The resulting models supportseveral additional languages while understanding images and executing tool calls.Inpublic benchmarks and human evaluations,both the server model and the on-devicemodel match or surpass comparably sized open baselines.A new Swift-centric Foundation Mod

5、els framework exposes guided generation,constrained tool calling,and LoRA adapter fine-tuning,allowing developers to inte-grate these capabilities with a few lines of code.The latest advancements in AppleIntelligence models are grounded in our Responsible AI approach with safeguardslike content filt

6、ering and locale-specific evaluation,as well as our commitment toprotecting our users privacy with innovations like Private Cloud Compute.1IntroductionApple Intelligence integrates powerful generative AI right into the apps and experiencesusers turn to everyday,all while protecting their privacy.At