Aya 23

What it is

Aya 23 is a family of multilingual instruction-tuned language models released with open weights by Cohere For AI ^[1]. It came in two sizes, 8 billion and 35 billion parameters ^[1]. A parameter is one of the numbers a neural network learns during training, and the count is a rough measure of how much the model can represent. "Instruction-tuned" means the model has been trained further on examples of a request paired with a good answer, so that it responds to instructions instead of merely continuing whatever text it is given.

"Open weights" means the trained parameters can be downloaded and run by others, rather than the model being reachable only through a company's interface. What people may do with those weights is set by a licence, covered below.

Where it came from

The model was introduced in a report posted to arXiv on 23 May 2024, with a second revision on 31 May 2024 ^[1]. It followed Aya 101, an earlier effort covering 101 languages. Aya 23 covers 23 languages, a deliberate narrowing: the team traded the breadth of Aya 101 for more capacity per language ^[1]. The reasoning is easy to state. Instruction data and model capacity are both finite; spread over a hundred languages, each language gets a thin slice, and spread over twenty-three, each gets more. A model that is only adequate everywhere may be less useful than one that is good in the languages a user actually needs.

In construction, Aya 23 pairs Cohere's pre-trained Command family of models with the Aya Collection instruction data ^[2]. In other words, the base models were not built from scratch for this release; existing multilingual base models were adapted with a multilingual instruction dataset.

What the authors report

The authors report that Aya 23 outperforms Aya 101 on the languages it covers ^[1]. They also report that it outperforms Gemma, Mistral and Mixtral on a range of discriminative and generative tasks ^[1]. The distinction matters: a discriminative task asks a model to choose or label — pick the correct answer among options, judge whether two sentences mean the same thing — while a generative task asks it to produce text, such as an answer, a summary or a translation. Multilingual models are often measured mainly on the second kind, so reporting both gives a fuller picture of ability. These are the authors' own reported comparisons, made at the time of release.

The 8-billion-parameter version has a context length of 8,192 tokens ^[2]. Context length is the amount of text a model can hold in view at once while answering; a longer context lets longer documents or multi-turn exchanges be handled without dropping earlier material.

Licence and access

The weights are released under the non-commercial CC-BY-NC 4.0 licence ^[2]. That licence permits reuse and modification with attribution but not commercial use. Cohere Labs' model card adds that using the model under this licence also requires adhering to Cohere Labs' Acceptable Use Policy ^[2], a separate document listing prohibited uses. For anyone planning to build a product, the licence is as important as the benchmark numbers: an openly downloadable model under non-commercial terms can be studied, reproduced and experimented with, but not simply shipped.

Where it sits

Aya 23 was succeeded in October 2024 by Aya Expanse, which kept the 23-language focus ^[2]. That continuity is itself informative about the project's direction: rather than returning to the wide-and-shallow coverage of Aya 101, the later release stayed with the narrower set of languages and worked on quality within it. Aya 23 therefore marks the point where the Aya line moved from "as many languages as possible" toward "fewer languages, done better."