AI startup Cohere on Tuesday released a new family of small AI models, Tiny Aya, to address the need for more diverse, multilingual models in the AI market.
tiny aya is one open weight model Cohere said this helps researchers, developers and communities create AI technology that is more reflective of their language and cultural contexts. TinyAya-Base is a 3.35B-parameter model covering over 70 languages. tinyia-global is an instruction-tuned multilingual model built on top of the base model.
the fog said model tokenizerOr a translation layer that converts text into a language that models understand, reducing segmentation across different linguistic structures, requiring fewer tokens per sentence across all languages. This reduces the need for additional tokens, thereby improving the inference efficiency of the model. Cohere also introduced special versions of Tinyaya, such as Tinyaya-Earth, which is strong throughout Africa and West Asia; tinea-fire, which is effective for South Asian languages; and is designed for languages in Tinian-Water, the Asia-Pacific region, and Europe.
Tiny Aya’s release reflects linguistic trends in the AI market and culturally diverse models. Many vendors are working to create models that can understand different languages and their nuances and respond in culturally relevant ways. For example, earlier this year, Google released TranslateGemma. However, translation models often forget the cultural context as most are trained in Chinese or English.
a need in the market
“This is a big deal because if you’re not able to train on the subtleties of a language, you have a training data set that doesn’t represent some of the subtleties that exist in that language,” said Bradley Shimmin, an analyst at Futurum Group.
He added that, because Tiny Aya is so young, CoHair is providing access to models in different languages democratizing technology itself, because the model is small enough to run on edge devices in remote locations.
“It can do the translation work for you without calling the cloud API, which is nice,” Shimin said. “I applaud them for doing that, and I think it’s certainly a necessary area of investment that, with a few exceptions, we haven’t been able to do much about.”
some challenges
However, the Tiny Aya’s small size could also mean limited use cases, said Mark Becque, an analyst at Omdia, a division of Informa TechTarget.
“I would love to know what they think is the best fit for the use cases,” Becque said. “What’s the market for this thing that they’re offering? Is it the American market? Why would people buy it?”
He said if Coher has designs on other regions outside the US, it must be sovereign AI-compliant, meaning the data it uses must be local to the region.
“They have to have a local presence to do these things,” Becque said.
