Apple’s open-source OpenELM small AI models released that could pave the way for AI features on the iPhone

Apple released its new family of artificial intelligence (AI) models called OpenELM. Short for Open Source Effective Language Models, there are a total of eight AI models with four pre-trained variants and four instruction variants. These are all small language models (SLMs) that specialize in text-related tasks, highlighting a fit with the tech giant’s reported ambitions to bring AI features to the device this year. Notably, the company is also said to have acquired a French AI startup called Datakalab, which works with computer vision models.

The OpenELM AI models were noticed on Apple’s Hugging Face page. Introducing SLM, the company said: “We are introducing OpenELM, a family of efficient open source language models. OpenELM uses a layer-wise scaling strategy to efficiently distribute parameters in each layer of the transformer model, resulting in increased accuracy.” There are two variants – pre-trained and instructed – and they come with 270 million, 450 million, 1.1 billion and 3 billion parameters.

Parameters refer to the neural networks in an AI model, which essentially means knowledge points, with each point in the network containing information about a specific topic. The greater the number of parameters, the more effectively it can understand and answer complex questions. For reference, the recently released Microsoft Phi-3-mini contains 3.8 billion parameters, while Google’s Gemma comes with 2 billion parameters. Pre-trained AI models are designed for general conversations and consistency in responses, while instruction variants are fine-tuned to perform a task.

Small language models may not exhibit a comprehensive knowledge base or conversational capacity like ChatGPT or Gemini, but they are efficient at handling specific tasks and requests and are generally less prone to errors. While Apple didn’t mention any specific use cases for the AI ​​models, it did offer the model weights available to the community. The weights are available under the Apple Sample Code License, which allows its use for both research and commercial purposes.

Apple, prone to SLM development, emphasizes that the company is focusing on its vision for on-device AI, as previously reported. The company has also published papers on three other AI models so far, including one that focuses on device capabilities, one that comes with multimodal capabilities, and another with computer vision that can understand smartphone screen interfaces.