OpenELM-270M

OpenELM-270M is the 270-million-parameter entry point of Apple's OpenELM family, released in April 2024. It uses a layer-wise scaling strategy to allocate parameters efficiently across transformer layers, and was pretrained on 1.8 trillion tokens of public data including RefinedWeb, PILE, RedPajama, and Dolma.

Apple published the full training and evaluation framework with checkpoints and logs, a first for the company. The weights carry the Apple Sample Code License, which is more restrictive than standard open-source licenses, so read it before commercial use.