AI startup Cohere on Tuesday released a new family of tiny AI models, Tiny Aya, to address a need for more diverse, multilingual models in the AI market.
Tiny Aya is an open weight model that helps researchers, developers and communities build AI technology that is more reflective of their language and cultural contexts, Cohere said. TinyAya-Base is a 3.35B-parameter model that covers over 70 languages. TinyAya-Global is an instruction-tuned multilingual model built on top of the base model.Â
Cohere said the model’s tokenizer, or translation layer that converts text into language that models understand, reduces breakup across different linguistic structures, thereby requiring fewer tokens per sentence across languages. This reduces the need for additional tokens, improving the model’s inference efficiency. Cohere also introduced specialized variants of TinyAya, such as TinyAya-Earth, which is strong across Africa and West Asia; TinyAya-Fire, which is effective for South Asian languages; and TinyAya-Water, designed for languages in the Asia-Pacific region and Europe.
The release of Tiny Aya reflects a trend in the AI market toward linguistically and culturally diverse models. Many vendors are working to create models that can both understand different languages and their nuances and respond in culturally relevant ways. For instance, earlier this year, Google released TranslateGemma. However, translation models often miss cultural context because most are trained on either Chinese or English.
A Need in the Market
“It’s a massive deal because if you’re not able to train on the subtleties of a language, then you end up with a training data set that does not represent some of the subtleties that might exist in that language,” said Bradley Shimmin, an analyst at Futurum Group.Â
He added that, because Tiny Aya is so small, Cohere is providing access to models in different languages and democratizing the technology itself, as the model is small enough to run on edge devices in remote locations.
“This can work on translations for you without having to call cloud APIs, which is nice,” Shimmin said. “I applaud them for doing this, and I feel like it’s definitely a needed area of investment that we’ve been without very much, with a few exceptions.”
Some Challenges
However, Tiny Aya’s small size could also mean limited use cases, said Mark Beccue, an analyst at Omdia, a division of Informa TechTarget.
“I’d love to know what they think the sweet spot is for use cases,” Beccue said. “What’s the market for this thing that they’re offering? Is it a US market? What are people going to buy this for?”
He added that if Cohere has designs on other regions outside the U.S., it must be sovereign AI-compliant, meaning the data it uses must be local to the area.
“They have to have a local presence to do these things,” Beccue said