Cohere Releases Tiny Aya: Open AI Models for 70+ Languages

Cohere, an enterprise AI company, launched a new set of open multilingual models called Tiny Aya at the India AI Summit. The models support over 70 languages, run on regular laptops without an internet connection, and come from the company's research team, Cohere Labs. This release aims to make AI tools available to more people around the world by focusing on languages often overlooked by big tech.

Background

Cohere has worked on multilingual AI for years. Their earlier projects include Aya 101, a model that handled 101 languages, and Aya 23, which covered 23 languages in detail. Those efforts came from Cohere for AI, a research group tied to the company. They built large datasets with help from thousands of researchers worldwide, including people from communities with less common languages.

The push for these models started because most AI tools focus on English or a few major languages. This leaves out billions of speakers in places like South Asia, Africa, and parts of Europe. Cohere gathered data in stages: first, broad training on raw text, then fine-tuning with specific instructions in many languages. For low-resource languages, they curated human-made examples and translated others to fill gaps.

Tiny Aya builds on this work but shrinks the size. The base model has 3.35 billion parameters, a count that measures how complex the AI is. This makes it light enough for everyday devices. Past models like Aya Expanse went up to 32 billion parameters and beat rivals from Google, Mistral, and Meta in tests across languages.

The India AI Summit provided the stage for this launch. Event leaders and AI experts gathered there to discuss global tech access. Cohere chose this spot to highlight support for South Asian tongues like Hindi, Bengali, and Tamil.

Key Details

The Tiny Aya family includes a base model and specialized versions. TinyAya-Global is tuned to follow user instructions better, ideal for apps needing wide language coverage. Then there are regional ones: TinyAya-Earth for African languages, TinyAya-Fire for South Asian ones, and TinyAya-Water for Asia Pacific, West Asia, and Europe.

Model Specs and Access

All models are open-weight, so developers can download, use, and change the code freely. The 3.35 billion parameter size lets them run locally, cutting costs and privacy risks from cloud services. They handle tasks like translation, summarization, and question answering.

South Asian support stands out with languages such as Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi. African and other regional variants aim to capture local speech patterns and culture.

“This approach allows each model to develop stronger linguistic grounding and cultural nuance, creating systems that feel more natural and reliable for the communities they are meant to serve. At the same time, all Tiny Aya models retain broad multilingual coverage, making them flexible starting points for further adaptation and research.” – Cohere statement

Developers can grab these from platforms like Hugging Face. Cohere Labs plans to share more data and tools to help others build on them.

What This Means

These models lower barriers for AI in non-English areas. Small teams or startups without big servers can now build apps in local languages. For example, education tools in rural India or health chatbots in Africa become possible without heavy costs.

Businesses gain from this too. Enterprises can add multilingual features to customer service or content tools without relying on closed systems. The open nature invites community tweaks, speeding up improvements for specific needs.

On a broader scale, Tiny Aya pushes the field toward equal language treatment. Past AI favored high-resource languages, but these models show small sizes can still perform well across dozens of tongues. Tests from earlier Aya versions already topped benchmarks in understanding and generation tasks.

Researchers now have new baselines. They can study how regional fine-tuning boosts accuracy or test on real-world data. This could lead to better handling of dialects or slang not in standard training sets.

For everyday users, the no-internet run means reliable access in spots with poor connections. Think farmers checking crop info in Punjabi or students learning in Swahili, all offline.

Cohere's move fits a trend where companies share weights to build trust and grow the ecosystem. It also counters worries about AI giants dominating non-English spaces. With over 70 languages covered, Tiny Aya helps bridge that divide.

The launch ties into ongoing work. Cohere keeps updating datasets and models, drawing from global input. Future versions might add more languages or shrink further for phones. For now, this family marks a step toward AI that speaks to everyone.