Apple has moved cautiously while much of the technology industry has rushed to put generative AI into products. That restraint is now looking less like absence and more like preparation.
A research paper posted online last Friday by Apple engineers describes MM1, a generative AI model built to work with both text and images. The paper does not prove what Apple will ship, but it gives a clearer view of the technical direction behind the company’s promised AI plans.
MM1 Shows Apple Is Building Multimodal AI
MM1 is described as a multimodal large language model, or MLLM. In practical terms, that means it can respond to text prompts while also interpreting images.
The model can answer questions about photos and show general knowledge abilities associated with chatbots like ChatGPT. Its name is not explained in the paper, though it could stand for MultiModal 1.
One example from the Apple research paper shows MM1 looking at a restaurant table with beer bottles and an image of a menu. When asked how much someone would expect to pay for all the beer on the table, the model reads the relevant price and calculates the total.
That kind of task matters because it combines several skills at once:
- Understanding a user’s written request.
- Identifying objects in an image.
- Reading information from another image.
- Using the visual and text context together to answer a practical question.
This is the direction many AI systems are moving. ChatGPT launched in November 2022 as a text-only system, while OpenAI and others later expanded large language model technology to work with more kinds of data. Google also emphasized Gemini’s multimodal nature when it launched the model last December.
Why The Paper Matters
Apple has not yet introduced even an AI-generated emoji, while rivals have pushed generative AI features more visibly. But MM1 suggests the company is making significant investments in the underlying technology.
The model appears similar in design and sophistication to recent AI models from other large technology companies, including Meta’s open source Llama 2 and Google’s Gemini. Work by rivals and academics suggests models of this type can power capable chatbots or help build agents that write code and act through computer interfaces or websites.
That does not mean MM1 will appear directly in an Apple product. Kate Saenko, a professor at Boston University who specializes in computer vision and machine learning, says it is hard to draw too many firm conclusions from the paper. Multimodal models can be adapted to many uses.
Still, Saenko suggests the research could point toward a multimodal assistant able to describe photos, documents, or charts and answer questions about them. For Apple, that kind of capability would naturally connect to the iPhone, where the camera, documents, messaging, and assistant features already sit close together.
Small Model, Big Strategic Signal
MM1 is described as a relatively small model based on its number of parameters, the internal values adjusted during training. Saenko says that smaller scale could make it easier for Apple’s engineers to test training methods and refinements before scaling up when they find something promising.
The paper also includes an unusual amount of training detail for a corporate AI publication. The engineers describe techniques used to improve the model, including increasing image resolution and mixing text and image data.
That openness stands out because Apple is known for secrecy. At the same time, the company has previously shown unusual openness about AI research as it tries to attract talent for a crucial field.
The timing also matters. Apple CEO Tim Cook has promised investors that the company will reveal more of its generative AI plans this year. The company is under pressure from rival smartphone makers, including Samsung and Google, which have already introduced generative AI tools for their devices.
The Siri Question
The iPhone already has an AI assistant: Siri. But the rise of ChatGPT and competing systems has made Siri look increasingly limited and out-of-date.
Amazon and Google have said they are integrating LLM technology into Alexa and Google Assistant. Google also lets Android phone users replace Assistant with Gemini.
Reports from The New York Times and Bloomberg say Apple is in preliminary talks with Google about adding Gemini to iPhones. That would echo Apple’s existing search strategy. Instead of building web search in-house, Apple relies on Google, which reportedly pays more than $18 billion to make its search engine the default on the iPhone.
Apple has also shown that it can move from an outside service to its own alternative. Google Maps used to be the default on iPhones, but in 2012 Apple replaced it with its own maps app.
That history leaves room for more than one path. Apple could use Gemini in some role while also developing in-house generative AI tools from MM1 and other models. Last September, several MM1 researchers published details of MGIE, a tool that uses generative AI to manipulate images based on a text prompt.
On-Device AI Could Be Apple’s Angle
Ruslan Salakhutdinov, a professor at Carnegie Mellon who led AI research at Apple several years ago, believes Apple may focus on LLMs that can be installed and run securely on Apple devices. That would match Apple’s emphasis on on-device algorithms, which can help protect sensitive data by avoiding unnecessary sharing with other companies.
Recent Apple AI research papers also concern machine-learning methods designed to preserve user privacy. This suggests privacy could be a key part of how Apple differentiates its generative AI work.
Apple may also have an advantage because it controls both software and hardware. The company has included a custom neural engine in the chips that power its mobile devices since 2017, starting with the iPhone X.
For now, MM1 is a research signal, not a product announcement. But it shows Apple has the expertise and infrastructure to train modern multimodal models. Brandon McKinzie, the Apple researcher who is lead author of the MM1 paper, wrote on X: “This is just the beginning. The team is already hard at work on the next generation of models.”