appleAt yesterday's WWDC24, it was announced Apple Intelligence(Apple Smart) will be iPhone, Mac and other devices introduce a series of AI functions.
Subsequently, Apple's machine learning official website released detailed information about Apple Intelligence. According to Apple's official introduction, Apple Intelligence has two basic models:
-
Local Model:On-device language model with approximately 3 billion parameters, the test score is higher than many open source models with 7 billion parameters (Mistral-7B or Gemma-7B);
-
Cloud Model: A larger cloud-based language model that can be run on Apple chip servers through private cloud computing.
Apple said that Apple Intelligence consists of multiple high-performance generative models that are specifically tailored to users' daily tasks and can dynamically adapt to their current activities. The basic models built into Apple Intelligence are fine-tuned for user experience, such as writing and refining text, prioritizing and summarizing notifications, creating interesting images for users' conversations with family and friends, and taking in-app actions to simplify interactions between apps.
In terms of pre-training, Apple's base model is trained on the AXLearn framework, an open source project released by Apple in 2023. It is built on JAX and XLA, enabling Apple to train models scalably on a variety of training hardware and cloud platforms, including TPUs as well as cloud and local GPUs.
IT Home noted that Apple promised that when training the basic model,The company never uses users' private personal data or user interactions and uses filters to remove personally identifiable information that is publicly available on the Internet., such as Social Security and credit card numbers. Apple also filtered profanity and other low-quality content to prevent it from being included in the training corpus. In addition to filtering, Apple also performed data extraction, deduplication, and applied model-based classifiers to identify high-quality documents.
In terms of optimization, Apple uses grouped-query-attention on both the device-side model and the server-side model. The on-device model uses a vocabulary size of 49K, while the server model uses a vocabulary size of 100K, which includes additional language and technology tags.
Through optimization, Apple claims to iPhone 15 Pro superior,Able to achieve a first token latency of approximately 0.6 milliseconds per prompt token and a generation rate of 30 tokens per second.
In the instruction trace evaluation (IFEval) test, Apple's local model performed better than models including Phi-3-mini, Mistral-7B and Gemma-7B, and was not inferior to DBRX-Instruct, Mixtral-8x22B and GPT-3.5-Turbo; while the level of the cloud model was basically on par with GPT-4-Turbo.
Apple plans to launch this summer iOS 18. Apple Intelligence will be opened in the iPadOS 18 and macOS Sequoia beta versions, and will then be open to the public in the form of beta versions this fall, but some features, more languages, and platform support will have to wait until next year.
Apple Intelligence is free to use, but is limited to devices with an A17 Pro chip or any M-series chip. This means that to use these features, you need an iPhone 15 Pro or iPhone 15 Pro Max, and the upcoming iPhone 16 series will also support Apple Intelligence.
On the Mac side, you need a Mac with an M1 or later, and for iPad, you need an iPad Pro or iPad Air with an M1 chip or later.