Apple releases OpenELM family of AI models for small on-device tasks: All you need to know

Apple’s AI push has been sluggish to say the least, especially when you compare it to the rapid developments taking place at its competitors, namely Microsoft and Google. Whilst the likes of Samsung, Google and even Nothing have a plethora of AI features on their respective devices, iPhones have remained aloof as Apple has been playing catch-up in the AI race. However, it is actively trying to make strides and has recently been in talks with the likes of Google and OpenAI over a possible deal that would allow their AI models to be used on iPhones, but this is still in development.

Now, Apple researchers have released a family of four lightweight AI models on the Hugging Face model library that can run on-device, hinting at their future use on devices such as iPhone, iPad and Mac.

Apple releases four open-source AI models

According to the post on Hugging Face, the family of AI models is known as ‘Open-source Efficient Language Models’ or OpenELM. These models have been designed to carry out small tasks efficiently, such as composing emails. Apple says that OpenELM is trained on publicly available datasets using the CoreNet library which includes RefinedWeb, deduplicated PILE, a subset of RedPajama, and a subset of Dolma v1.6, totalling approximately 1.8 trillion tokens. It has been released with four parameters - 70 million, 450 million, 1.1 billion, and 3 billion parameters.

For the unaware, parameters are a measurement of how many variables the AI model can learn from while making decisions. These are based on the dataset the AI model has been trained on.

According to Apple, the OpenELM family of AI models has been released to “empower and enrich the open research community by providing access to state-of-the-art language models”.

Apple’s AI push

The iPhone maker has been experimenting with AI for some time now. Last year, it released a machine learning framework called MLX that enables AI models to run better on its devices powered by Apple Silicon. Moreover, it also launched an image tool called MLLM-Guided Image Editing or MGIE.

Last month, it was revealed that Apple researchers have had a breakthrough when it comes to training AI models on both text and images. A research paper on the same was published on March 14. Titled “MM1: Methods, Analysis & Insights from Multimodal LLM Pre-training”, it demonstrates how using several architectures for training data and models may help achieve state-of-the-art results across multiple benchmarks.

It is also said to be working on its own Large Language Model (LLM), at the heart of which is a new framework known as Ajax that could bring a ChatGPT-like app, nicknamed "Apple GPT”. Collaboration across various departments at Apple, such as software engineering, machine learning, and cloud engineering, is said to be taking place to make this LLM project a reality.

The release of the OpenELM family of AI models certainly paints an intriguing picture of the AI development at Apple. However, considering that no foundational model has been released yet, there is some time before Apple devices such as iPhone and Mac will finally be able to take advantage of it.