AI enterprise Cohere unveiled a novel collection of polyglot models concurrent with the India AI Summit. These models, christened Tiny Aya, possess open-weight status — signifying their foundational programming is openly accessible for utilization and alteration by all — accommodate more than 70 tongues, and function on common gadgets such as portable computers, even offline.
This model, introduced by Cohere Labs, the firm’s investigative division, provides compatibility for South Asian tongues, including Bengali, Hindi, Punjabi, Urdu, Gujarati, Tamil, Telugu, and Marathi.
The foundational model encompasses 3.35 billion parameters, indicating its scale and intricacy. Additionally, Cohere introduced TinyAya-Global, an iteration refined for enhanced adherence to user instructions, catering to applications needing extensive linguistic coverage. Regional iterations complete the series: TinyAya-Earth addresses African tongues; TinyAya-Fire serves South Asian dialects; and TinyAya-Water covers Asia Pacific, West Asia, and Europe.
“Such a methodology enables every model to cultivate more robust linguistic foundations and cultural subtleties, thereby generating systems that appear more authentic and trustworthy for their intended user bases. Concurrently, every Tiny Aya model preserves extensive polyglot capability, positioning them as versatile initial platforms for subsequent customization and investigation,” the firm declared in an announcement.
Cohere highlighted that these models, having been developed utilizing a sole cluster of 64 H100 GPUs (Nvidia’s powerful processors) and comparatively limited computational resources, present an optimal solution for researchers and developers crafting applications for populations using indigenous tongues. These models possess the ability to operate directly on hardware, enabling developers to facilitate offline translation. The firm remarked that its foundational software was engineered for on-device deployment, demanding less processing capability than a majority of similar models.
Within nations boasting linguistic variety, such as India, this sort of disconnected functionality can unlock a varied array of programs and scenarios, bypassing the necessity for continuous web connectivity.
These models can be found on HuggingFace, a well-known hub for distributing and assessing AI models, as well as on the Cohere Platform. Programmers are able to obtain them from HuggingFace, Kaggle, and Ollama for deployment on local systems. The firm will additionally publish training and assessment datasets on HuggingFace and intends to issue a technical document outlining its training approach.
Techcrunch event
Boston, MA
|
June 23, 2026
Aidan Gomez, the fledgling company’s chief executive, indicated last year that the enterprise intends to list publicly “in the near future.” As reported by CNBC, the firm concluded 2025 robustly, achieving $240 million in yearly recurring income, alongside a 50% quarterly expansion across the entire year.
{content}
Source: {feed_title}