8 mini AI models for smartphones open up great opportunities

In the world of artificial intelligence, so-called “small language models” are gaining popularity, which can run on a local device instead of powerful cloud services. Apple recently introduced something interesting – an open source set of tiny AI programs called OpenELM. They are so compact that they can run directly on a smartphone.

Although at the moment OpenELM (Open-source Efficient Language Models) is just a research project, in the future it could become the foundation for completely new solutions from Apple. We are talking about local data processing technologies that will allow the company to provide the highest possible level of confidentiality and protection of personal data for clients.

The OpenELM source code is available on the popular Hugging Face platform under the Apple Sample Code License. Although this license contains some restrictions that prevent OpenELM from being considered a completely open source project in the conventional sense, the model files themselves are freely available.

Microsoft recently introduced Phi-3, a similar product with the same goal of achieving efficient natural language processing in a small local neural network. However, OpenELM turned out to be even more miniature.

Apple has released as many as eight different flavors of OpenELM. Their volume varies from a very modest 270 million parameters to 3 billion:

• OpenELM-270M

• OpenELM-450M

• OpenELM-1_1B

• OpenELM-3B

• OpenELM-270M-Instruct

• OpenELM-450M-Instruct

• OpenELM-1_1B-Instruct

• OpenELM-3B-Instruct

Four of them have the most basic functions. For example, they predict the next words in a text by analyzing previous sentences. The remaining four models have undergone more specialized tuning to understand and follow instructions from users. They are already much more suitable for use in interactive applications and chatbots.

All eight OpenELM models are capable of processing up to 2048 words at a time. This allows them to work with impressive amounts of text.

Compared to leading large language models like Meta’s Llama 3 with 70 billion parameters or OpenAI’s GPT-3 with 175 billion, Apple’s new products look truly tiny. However, this was the essence of recent research – to create algorithms that, with fewer settings, will not be inferior to the giants in functionality.

According to the developers, a key feature of their approach with OpenELM was the company’s “layered scaling” technique. It allows you to optimally distribute parameters across the layers of the neural network, achieving maximum efficiency.

This solution not only saves computing resources, but also improves performance when training on relatively small amounts of data. Thanks to its layered scaling technique, OpenELM models achieved 2.36% higher accuracy than Allen AI’s OLMo 1B while using half the number of tokens, according to Apple’s white paper.

Most importantly, Apple not only published the source code for the OpenELM models themselves, but also released the code for the CoreNet library that was used to train them. In addition, the company provided detailed training instructions that will allow the neural network weights to be replicated. This unprecedented level of transparency is still rare, even in developments from leading technology giants.

Apple has not yet integrated the latest developments into its consumer devices. However, the upcoming iOS 18 update, due to be unveiled in June at WWDC, is rumored to include new features with local processing to ensure user privacy. However, it is possible that for more complex tasks that require cloud computing, Apple may hire third-party companies like Google or OpenAI to finally improve the capabilities of the Siri voice assistant.