Apple continues AI innovation with new open-source language models

Apple's Apple Intelligence research team has unveiled two new high-performing language models designed to enhance AI training.

These models, part of the open-source DataComp for Language Models project, have proven competitive with industry leaders like Llama 3 and Gemma.

The models, one with seven billion parameters and a smaller one with 1.4 billion, are used to train AI systems such as ChatGPT by providing a framework that includes architecture, parameters, and filtered datasets. Apple reports that the larger model outperformed the previous top model, MAP-Neo, by 6.6 percent in benchmarks while using 40 percent less computing power.

These models are fully open-source, with the dataset, weight models, and training code available for researchers.

Apple's Machine Learning team has shared these innovations to contribute to community research rather than future Apple products.