Apple Just Released Details About Its New MM1 AI Model

We’ve already learned that Apple may tap Google’s Gemini to power some of iOS 18’s new AI features, but that hasn’t stopped the tech giant from working on its own AI models. In a new paper, Apple has revealed more details about how it is approaching the development of its new MM1 AI model.

Apple says it plans to use a diverse dataset that includes interleaved image-text documents, image-caption pairs, and text-only data to help train and develop MM1. This, Apple claims, should allow the MM1 model to set a new standard in AI’s ability to create image captions, answer visual questions, and even how it responds with natural language inference. The hope, it seems, is to provide as high a level of accuracy as possible.

This method of research allows Apple to focus on several types of training data and even architectures for the model, which should provide greater room for the AI to understand and even generate language based on both linguistic and visual cues. 

Apple clearly hopes that by combining training methods from other AI with its own methods, it will be able to provide better pre-training metrics and achieve a competitive performance that will help it catch up with other companies that are already deep into AI development, like Google and OpenAI.

Apple has never been a stranger to paving its own way. The company continues to find new ways to approach the same situations that other companies have, including in the way it designs its hardware and software. Whether you think that’s a good thing or not is up to you, but the matter stands that Apple’s ongoing attempts to create a reliable and competitive AI was always going to approach things in a different way, and based on the information showcased in this paper, the company has found a very unique way to do that.

This paper, of course, is just our first real look into the things Apple is doing to build out its AI capabilities. It will be interesting to see where things go from here.