Cohere Unveils Open-Weight Tiny Aya Multilingual AI Models for On-Device Applications at India AI Summit.

Enterprise artificial intelligence powerhouse Cohere has announced the launch of an innovative new family of multilingual models, dubbed Tiny Aya, on the sidelines of the ongoing India AI Summit. These groundbreaking models are characterized by their open-weight nature, signifying that their foundational code and architecture are publicly accessible, allowing developers, researchers, and anyone interested to freely use, inspect, and modify them. This commitment to openness is poised to foster collaborative innovation within the global AI community. A key highlight of the Tiny Aya family is its extensive linguistic coverage, supporting more than 70 languages, thereby addressing a significant need for broader linguistic inclusivity in AI. Furthermore, these models are engineered for efficiency, capable of running directly on everyday devices such as laptops without the necessity of an active internet connection, marking a substantial leap towards accessible and decentralized AI processing.

The development and launch of the Tiny Aya models were spearheaded by Cohere Labs, the company’s dedicated research arm, underscoring Cohere’s commitment to pushing the boundaries of AI research and practical application. A particular focus for the Tiny Aya family is its robust support for a diverse array of South Asian languages. This includes widely spoken languages such as Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi. This targeted linguistic support is crucial for empowering millions of users in a region known for its rich linguistic diversity, enabling them to interact with AI technologies in their native tongues and fostering greater digital inclusion.

At its core, the base Tiny Aya model is defined by its 3.35 billion parameters. In the realm of artificial intelligence, parameters are a measure of a model’s size and complexity, directly influencing its capacity to learn and perform tasks. While 3.35 billion parameters position Tiny Aya as a compact yet powerful model, it is designed to strike an optimal balance between performance and the ability to run efficiently on local devices. To cater to a broader spectrum of applications and user needs, Cohere has also introduced TinyAya-Global. This particular version has been meticulously fine-tuned to enhance its ability to interpret and follow user commands more effectively, making it an ideal choice for applications that demand extensive and reliable language support across various domains.

Beyond the global variant, the Tiny Aya family is further diversified into specialized regional models, each tailored to specific geographical and linguistic contexts. These regional variants exemplify Cohere’s strategic approach to providing culturally nuanced and linguistically precise AI solutions. TinyAya-Earth is specifically designed to cater to African languages, acknowledging the vast linguistic tapestry of the continent. TinyAya-Fire focuses on the intricate linguistic landscape of South Asian languages, building upon the core model’s strengths in this region. Completing the family, TinyAya-Water is engineered to serve the diverse language requirements of the Asia Pacific, West Asia, and European regions. This segmented approach ensures that each model can offer highly relevant and effective linguistic processing for its intended audience.

In a statement released by the company, Cohere emphasized the profound benefits of this specialized approach. "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," the company articulated. This philosophy highlights Cohere’s dedication not just to technical performance but also to creating AI that genuinely resonates with users on a cultural and linguistic level. The statement further noted, "At the same time, all Tiny Aya models retain broad multilingual coverage, making them flexible starting points for further adaptation and research." This balance between specialized nuance and broad applicability positions the Tiny Aya models as versatile tools for both immediate deployment and future AI innovation.

Cohere also provided insights into the training methodology and performance capabilities of these new models. Remarkably, the Tiny Aya models were trained on a single cluster comprising 64 Nvidia H100 GPUs. H100 GPUs are high-powered chips renowned for their exceptional processing capabilities in AI workloads. The fact that Cohere achieved this with a "relatively modest computing source" underscores the efficiency of their training methodologies and the optimized architecture of the Tiny Aya models. This efficiency makes them particularly well-suited for researchers and developers who are building applications specifically for audiences that communicate in native or less-resourced languages.

The capability of these models to run directly on devices is a game-changer, opening up a plethora of practical applications. Developers can leverage Tiny Aya to power offline translation services, enabling seamless communication even in environments without internet connectivity. This is a critical feature, especially in regions where internet access can be inconsistent or expensive. Cohere highlighted that its underlying software was specifically engineered to facilitate on-device usage, demanding significantly less computing power compared to most comparable large language models. This meticulous optimization ensures that Tiny Aya can deliver high performance without overburdening device resources, making it accessible on a wider range of hardware.

The implications of such offline-friendly capabilities are particularly profound in linguistically diverse countries like India. In a nation where hundreds of languages and dialects are spoken across various regions, the ability to deploy AI models locally and without constant internet access can unlock an unprecedented array of applications and use cases. This includes localized educational tools, accessible government services, enhanced communication platforms, and myriad other innovations that can bridge digital divides and empower communities by enabling technology to speak their language, anywhere, anytime.

For developers and researchers eager to integrate these models into their projects, Cohere has ensured broad accessibility. The Tiny Aya models are readily available on HuggingFace, a highly popular platform within the AI community known for sharing and testing machine learning models. They are also accessible via the official Cohere Platform. Furthermore, developers have multiple avenues for local deployment, with the models available for download on HuggingFace, Kaggle (another prominent platform for data science and machine learning), and Ollama, a lightweight tool for running open-source large language models locally. To further support the research community and promote transparency, Cohere is also releasing the comprehensive training and evaluation datasets used for Tiny Aya on HuggingFace. The company has also announced plans to release a detailed technical report outlining its innovative training methodology, offering valuable insights for future AI development.

This strategic product launch comes at a significant juncture for Cohere as a company. Aidan Gomez, the CEO of Cohere, indicated last year that the startup intends to go public "soon," signaling a period of anticipated growth and expansion. Recent financial reports underscore the company’s strong market position. According to a report by CNBC, Cohere concluded 2025 on a robust financial note, achieving an impressive $240 million in annual recurring revenue (ARR). This substantial revenue was supported by a remarkable 50% quarter-over-quarter growth throughout the year, demonstrating strong market adoption and demand for Cohere’s enterprise AI solutions. The introduction of the Tiny Aya family further solidifies Cohere’s commitment to advancing accessible, multilingual AI, aligning with its broader vision for an inclusive and powerful AI future.