Understanding Apple Intelligence, Apple’s great own model: this is how it looks compared to other AIs and what it will use ChatGPT for

Apple Intelligence Llm Model

Apple’s AI has not arrived as we thought. When the collaboration with OpenAI was leaked, we thought that Sam Altman’s company was going to have much more weight, but in the end it has not been so. Apple Intelligence is an artificial intelligence model entirely developed by Apple, with no relation to ChatGPT.

Table of Contents

The alliance between Apple and OpenAI is official, but we are looking at two separate projects. On the one hand, Apple Intelligence and Siri and on the other the addition of ChatGPT. Here we explain what Apple has developed, how OpenAI’s role is and how Apple Intelligence is positioned in this battle of artificial intelligence where the Cupertino company is already participating with its own weapons.

During the presentation of Apple Intelligence, its functions were announced, that it would be integrated into iOS 18, iPadOS 18 and macOS Sequoia and that it would have both an ‘on-device’ version and a version connected to the cloud, but privately with what they called ‘Private Cloud Compute’. But they did not give technical details about the AI model behind it.

Fortunately, Apple’s Machine Learning Research team has published a small report on the technical data of the language model (LLM) that Apple Intelligence moves. Here we review its main features, which will help us to better understand how far Apple’s AI that has been presented to us during WWDC 2024 goes.

A proprietary AI model with 3,000 million parameters

Apple Intelligence is a set of several AI models specialized in “daily tasks for users.” These models have been fine-tuned to excel at tasks such as writing, summarizing, creating images, and simplifying interactions between applications. That is, many of the tasks that we can perform from Siri on the iPhone.

As described by Apple, its AI model has about 3,000 million parameters for the version on the device and an even larger model that works on Apple servers that work with Apple Silicon chips and renewable energy. That is, an on-device AI model and another on Apple servers.

Apple is currently reserving these two AI models, but explains that it will unveil new models soon.

If we compare it to its direct rival, Apple Intelligence is at a practically identical level to Gemini Nano. If Apple’s offers 3,000 million parameters, Google’s Gemini Nano has two versions, one with 1,8000 million for devices with low RAM memory and another with 3,250 million parameters, for the Pixel and other high-end Android.

To train their model, they rely on Apple AXLearn, an open source project started in 2023. As for the content used to create this model, Apple explains that it has used licensed data, as well as publicly available content, collected by its AppleBot web crawler.

Apple points out that they never use users’ private data or interactions to create this AI model and that they apply filters to personally remove any information that could identify users and that is on the web. Other filters applied are to remove low-level or duplicate data. And finally, they explain that websites that do not want to see their content used for Apple Intelligence, have the option of requesting it. These are arguments that we have seen repeated on numerous occasions, which indicates that to create its AI, Apple has followed the traditional recipe.


AI training also has its respective optimization. Here Apple explains that they have reduced the memory usage needed. The on-device version has a vocabulary of 49K, while the server model has 100K. In comparison, Gemini Nano has a token limit of 32K.

As we explained initially, Apple’s model has been trained to perform specific tasks. Apple defines them as ‘adapters’. A set of specific parameters to improve the tasks they want to perform. For example, at the level of summarizing, Apple uses 750 specific questions to evaluate whether the answer is correct. Among the tasks in which they have applied additional work are those of “brainstorming, classification, answering closed questions, programming, extraction, mathematical reasoning, answering open questions, rewriting, security, summarizing and writing”.

On a par with GPT-4 Turbo, in specific tasks

Apple shows several benchmarks comparing themselves to some of the leading AI models of the moment. In the on-device AI model, it is compared to Microsoft’s Phi-3 mini, Google’s Mistral 7-B, and Gemma 7-B. While in the server model it is compared to GPT-4 Turbo and Mistral 8x22B.

Benchmarks Apple

According to the IFEval benchmark, which measures the ability of AI to follow instructions, Apple Intelligence models are on par with and even outperform open source and commercial models of equivalent size.

They also teach several benchmarks for writing and summarizing texts, where Apple’s AI especially stands out. However, none of Apple’s chosen benchmarks are generic. We will have to wait for future tests to determine to what extent Apple’s AI manages to compete against the main models or is only up to their level in the tasks for which it has been specially tuned.

Apple has adopted a specialization strategy for its AI, but we have no trace of multimodality. We do have text and also voice, but at no time is video referenced, as Google does with Project Astra or OpenAI with GPT-4o. It’s a shame, because future Apple Vision Pro could put a lot of use into it.

How Apple will protect information when it’s sent to the cloud

The promise of Apple Intelligence is this: AI functions are performed on the device and processed with the on-device AI model. But when the task is too complex, it will choose to connect to Apple’s servers.

The goal is for most tasks to be done directly on the iPhone, iPad, or Mac. To do this, Apple requires at least an iPhone 15 Pro (with the A17 Pro) or an M1 processor or higherWhen Apple’s processor is not powerful enough for that task we ask of AI, Private Cloud Compute will come into action. Some cases, for example, are the analysis of images, the summary of large texts or the search for information on the internet.

As Apple explains, they have created a specific cloud infrastructure for AI. One that they promise is the “most secure and advanced ever created for AI in the cloud.” And what do they do differently? Here are some of the promises Apple explains.

For example, the sending of data is end-to-end encrypted and when this is not possible, they “strive to process user data ephemerally or under uncorrelated random identifiers that obscure the user’s identity.”

Apple recognizes that security and privacy in the cloud are difficult to verify and secure. Once Private Cloud Compute is available in beta for developers they will release more details, but for the moment they promise that they will force certain requirements such as that “user data can only be used for the purpose requested by the user himself” and that none of this data will be visible, “not even by Apple employees”.

At the component level, Apple intends that no external element be used for security. And in the case of error analysis or server metrics, services with a high level of privacy will be used.

The role of Apple workers is relevant and they explain that Private Cloud Compute will not offer any additional permission to workers to bypass the restrictions. Finally, Apple explains that it will allow any security researcher to verify the operation of its infrastructure to verify that the guarantees are met.

Siri will connect with ChatGPT, but it won’t be the only one

Chatgpt Siri

As we can see, at no time in the development of Apple Intelligence does OpenAI intervene. Where is the alliance then? From the end of this year, ChatGPT will be integrated into Siri and writing tools.

For certain questions, Siri will refer the answer to ChatGPT and it will answer us directly. This means that Siri users will have free access to ChatGPT, without the need to create a new account.

There is a point that Apple has explained but not taught. When we want to use ChatGPT, Siri will send us a notification to accept. From accepting this use of ChatGPT we enter a different terrain. Once we give access to ChatGPT, our data will be transferred to OpenAI’s servers.

However, Apple promises that requests will not be able to be stored by OpenAI and that users’ IPs will be hidden, following the promise of Private Cloud Compute. This phase of data processing is the one that generates the most privacy concerns.

The partnership with OpenAI will not be exclusive. Craig Federighi, Apple’s VP of engineering, explained during the conference that in the future new AI systems may be added to Siri, beyond ChatGPT. Without ruling out even Google Gemini. With this move, Apple has an ace up its sleeve to find new allies; continue to develop their own AI; not offering a closed system to regulators and being able to cover markets where ChatGPT is not available.

This is the case of China, a market that Apple aspires to and where iPhones are very popular. With this Apple Intelligence strategy, Siri could well work alongside the AI systems of companies like Baidu or even a state-run AI. Apple has created its own AI model, but it has also opened the door for other companies to help it reach where they can’t reach.




Leave a Reply