Cohere For AI releases Aya 23 open-weight models

Background

Cohere For AI published Aya 23 as a report posted to arXiv. The work is a multilingual model family, and the work is by Cohere For AI ^[1].

What happened

On 23 May 2024 Cohere For AI published Aya 23, a family of multilingual models covering 23 languages ^[1]. Open weights were released in 8B and 35B sizes ^[1]. The models build on Cohere's pre-trained Command family ^[2].

Key details

Unlike the earlier Aya 101, Aya 23 focuses on depth over breadth, devoting more capacity to fewer languages ^[1]. The team reported that it beats Aya 101, Gemma, Mistral and Mixtral across many tasks ^[1].

The weights carry a non-commercial licence ^[2]. Anyone who wants to use the models commercially therefore needs a different arrangement.

Impact

Aya 23 kept the multilingual line focused on a fixed set of 23 languages, a choice the series kept in its successor, Aya Expanse ^[2].