Apple Releases On-Device AI Model OpenELM

Apple has recently announced a new open-source LLM (Large Language Models) known as OpenELM (Open-source Efficient Language Models).

Apple’s LLM will run locally on devices, making a great shift from costly cloud-based processing to on-device processing. Apple’s main focus for this LLM is to improve processing speed with enhanced privacy.

OpenELM models use a layer-wise scaling technique, that effectively allocates parameters with every layer of the transformer model to increase accuracy. For example, with a budget of around one billion parameters, OpenELM achieved a 2.36% gain in accuracy over its predecessor model, OLMo but using half the amount of pre-training tokens.

This method not only increases speed but also decreases the computational burden on devices, which is critical for running AI applications directly on consumer hardware.

Features and Capabilities

The OpenELM project has several major elements that set it apart from past AI models:

• Open-source availability: Apple has made OpenELM available on the Hugging Face Hub, allowing developers and researchers to access and participate in its development.

• Comprehensive Training Framework: Unlike traditional models, which simply supply model weights and inference code, the OpenELM release contains the entire infrastructure for training and evaluating publicly available datasets. This features training records, many milestones, and pre-training setups.

• Enhanced privacy and speed: Because OpenELM runs on-device, there is no need to send data to cloud servers, which improves user privacy. Furthermore, local processing minimizes latency, resulting in quicker reaction times for AI-powered features on device
.

Integration with iOS

Apple intends to include OpenELM in the future iOS 18 version, which is likely to include several new AI features. The integration of OpenELM is anticipated to underpin a variety of on-device AI features, including more powerful versions of Siri and other AI-powered applications.