Multilingual models are entering a new phase of innovation, and today Cohere is introducing Tiny Aya — a new type of AI system designed to understand and generate text across many languages. The goal of this launch is to make language technology easier to access, more efficient, and inclusive around the world.
Tiny Aya is a big improvement for Multilingual Models. It translates well, understands different languages and can write texts — and all in one small program that you can use on your computer or mobile phone.
You don't need to spend money on complicated equipment or software. There are no major technical problems. It's just an efficient, capable AI model that can work almost anywhere.
Rethinking Multilingual AI
The technology that allows computers to understand and respond to more than one language has got much better very quickly over the past few years. But most of the time, the performance is in a few popular languages and big cloud setups.
Tiny Aya is different. Instead of supporting lots of different languages, it focuses on supporting a few languages well, making sure they work well in different regions, including less well-known languages.
This approach lets researchers, developers and communities build AI systems that reflect their own languages and cultures, not just global internet trends.
What Cohere Is Releasing
As part of this big update to Multilingual Models, Cohere introduced Tiny Aya along with a wider range of open multilingual systems and supporting research resources.
The website TechCrunch reported this development, and said that Cohere is launching open multilingual models designed to improve language accessibility while maintaining strong real-world performance.
TinyAya-Base (3.35B Parameters) is a pretrained multilingual foundation model supporting 70+ languages, including many that are not often used in AI systems.
TinyAya-Base (3.35B Parameters)
TinyAya-Base is a pretrained multilingual foundation model supporting 70+ languages, including many that are not often used in AI systems. This wide range of coverage makes it a very useful tool for research, fine-tuning, and local use in different regions.
The supported languages include:
The languages include Amharic, Arabic, Basque, Bengali, Bulgarian, Burmese, Cantonese, Catalan, Chinese (Simplified and Traditional), Croatian, Czech, Danish, Dutch, English, Estonian, Finnish, French, Galician, German, Greek, Gujarati, Hausa, Hebrew, Hindi, Hungarian, Igbo, Indonesian, Irish, Italian, Japanese, Javanese, Khmer, Korean, Lao, Latvian, Lithuanian, Malagasy, Malay, Maltese, Marathi, Nepali, Nigerian Pidgin, Norwegian (Bokmål), Persian, Polish, Portuguese, Punjabi, Romanian, Russian, Serbian, Shona, Slovak, Slovenian, Spanish, Swahili, Swedish, Tagalog, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu, Vietnamese, Welsh, Wolof, Xhosa, Yoruba, and Zulu.
TinyAya-Base covers a wide range of languages, from well-known global languages like English, Spanish and Mandarin Chinese to less well-known languages like Wolof, Yoruba and Shona. This makes modern Multilingual Models more inclusive.
TinyAya-Global
TinyAya-Global is built on TinyAya-Base. It is an instruction-tuned model that is optimised for consistent performance across 67 supported languages.
It is perfect for companies that need:
- The performance is balanced across all languages
- Use one model in different areas
- Good at following instructions
This makes it a practical choice for systems that use more than one language, like customer support, education, and business AI.
Specialized Regional Models
As well as the global model, Cohere is releasing region-focused versions that improve performance in specific language ecosystems while keeping good cross-lingual ability.
These include:
- TinyAya-Earth is the strongest for Africa and West Asia.
- TinyAya-Fire is best for South Asian languages.
- TinyAya-Water is the strongest in Asia-Pacific and Europe.
This family of builders combines shared multilingual learning with regional specialisation. This means that builders can choose models based on their target communities.
New Multilingual Datasets and Benchmarks
Tiny Aya is supported by a large multilingual fine-tuning dataset covering many areas and tasks.
These resources enable:
- You can repeat the evaluation
- Experimenting with data
- Keep working on training strategies for people who speak more than one language.
Cohere is also releasing a detailed report. This will explain the training strategy, evaluation methods and design principles behind Tiny Aya.
Performance Across Languages
Tiny Aya is as good as other massively multilingual systems.
Across different tests covering:
- Translation
- Understanding what someone is saying
- Using maths to solve problems
- Generation that goes on forever
Tiny Aya shows the latest technology for creating music in different languages, especially from West Asia and Africa.
Unlike many other large language models, it performs well even for languages that are not used very often on the web.
Technical Innovation Behind Tiny Aya
Tiny Aya is based on years of research into many languages from the Aya project.
The most important new ideas are:
Advanced Tokenization
The tokenizer reduces fragmentation across different languages and scripts, producing fewer tokens per sentence. This makes things more efficient and reduces the amount of computing power needed.
Improved Language Plasticity
Research-driven methods improve how well computers can adapt to different languages, while also keeping the details of the language.
Smarter Synthetic Data Integration
By making synthetic data more natural and merging different generations carefully, the model keeps both cross-lingual strength and language-specific accuracy.
Efficient Post-Training
Tiny Aya's post-training was completed on a single 64 NVIDIA H100 GPU cluster — showing that clever data design and training strategies can take the place of simply using more powerful computers.
This efficiency makes Tiny Aya suitable for use in the real world without losing any quality.
Small Models, Strong Performance
Tiny Aya shows that how big a model is doesn't decide what it can do.
Even though it is small, it performs well across different languages, especially in:
- How well the translation is accurate
- Using maths to solve problems
- Generation quality that will always be good.
- Tasks involving language that do not require many resources
Instead of trying to get the highest scores on tests, Cohere focused on making sure the software was stable, easy to use, and useful for real-world situations — things that are important to native speakers.
Accessibility as a Core Design Principle
The way Tiny Aya is designed is all about making it easy for people to use.
The model is small enough to be used in classrooms, community labs, and regions with limited cloud infrastructure. For example, a university lab in India could use Tiny Aya to translate or help with AI without using other websites.
By making it easier for people to experiment with multilingual AI, Cohere helps more communities to innovate and shape this technology in a way that works for them.
A New Ecosystem of Multilingual Models
Tiny Aya shows how Multilingual Models are made and shared in a different way.
Instead of one massive, centralized AI system, the future will have a distributed ecosystem of specialised models — each adapted to local linguistic needs and research priorities.
Cohere invites researchers and developers worldwide to improve and experiment with these models, which will help to create a more diverse and inclusive AI landscape.
Final Thoughts
Cohere's Tiny Aya is more than just a model release. It represents a major step forward in the evolution of Multilingual Models and signals a powerful shift in AI language processing. Rather than focusing only on scale, it emphasizes inclusion, balance, and real-world usability.
As Multilingual Models continue to advance, systems like Tiny Aya demonstrate that strong performance does not always require massive datasets or oversized architectures. Instead, success comes from thoughtful design, responsible research, and a genuine commitment to supporting linguistic diversity. This new generation of Multilingual Models proves that AI can be both powerful and inclusive at the same time.
FAQs
1. What are Multilingual Models in AI?
Multilingual Models are AI systems designed to understand, process, and generate text in multiple languages using a single unified architecture. They learn shared linguistic patterns across languages, making them more efficient than building separate models for each language.
2. What is TinyAya-Base by Cohere?
TinyAya-Base is a 3.35B-parameter open multilingual foundation model developed by Cohere. It supports over 70 languages and is optimized for efficiency, making it suitable for research, education, and enterprise AI applications.
3. Why are Multilingual Models important for global businesses?
Multilingual Models allow businesses to provide customer support, translation, and AI-driven communication across multiple languages using one system. This reduces infrastructure costs and improves accessibility in international markets.
4. How many languages do Cohere’s Multilingual Models support?
Cohere’s TinyAya-Base supports 70+ languages, including major global languages and underrepresented regional languages, helping bridge the digital language gap.
5. Are Multilingual Models better than single-language AI models?
For global applications, yes. Multilingual Models streamline development, improve scalability, and enable cross-language understanding, making them ideal for international platforms and diverse user bases.
Sign in to leave a comment.