Skip to main content
Back

Cohere Labs' Tiny Aya Brings 70-Language AI to Laptops Without Internet

Cohere Labs releases tiny-aya-global, a 3.35B open-weight model covering 70+ languages that runs locally on a laptop or phone

  • Open weights release: tiny-aya-global is a 3.35B instruction-tuned model covering 70+ languages, free to use for non-commercial purposes.
  • Runs locally: Small enough to run on a laptop or phone without an internet connection, no cloud required.
  • Four regional variants: Global (balanced), Earth (Africa/West Asia), Fire (South Asia), and Water (Asia-Pacific/Europe) for targeted language performance.
  • Hybrid attention architecture: Combines sliding window local attention with a global attention layer, keeping inference efficient while handling long-range context.
  • Best for generation tasks: Strong on translation, summarization, and conversational AI; weaker on chain-of-thought math reasoning.
  • Available now: Via HuggingFace Space, Cohere API, Transformers, vLLM, and Docker under CC-BY-NC license.

Most capable multilingual models are either massive, English-heavy, or both. tiny-aya-global is Cohere Labs' answer to that gap: a 3.35 billion parameter, instruction-tuned model that covers 70+ languages and is small enough to run on a laptop without an internet connection. It is part of the broader Tiny Aya family, and it is available now as open weights on Hugging Face.

The problem with multilingual AI today

Most translation and multilingual models miss cultural context because they are trained primarily on English or Chinese data. The result is that billions of speakers of languages like Swahili, Yoruba, Bengali, or Javanese get AI that technically understands their language but fumbles the nuance. As one analyst put it, "if you're not able to train on the subtleties of a language, then you end up with a training data set that does not represent some of the subtleties that might exist in that language."

Tiny Aya represents an alternative scaling path for multilingual AI: one centered on efficiency, balanced performance across languages, and practical deployment rather than simply increasing parameters. The bet is that a well-trained 3B model, built from the ground up for multilingual coverage, can outperform much larger general-purpose models on the languages that matter most to underserved communities.

What tiny-aya-global actually is

Tiny Aya is an open weights research release of a pretrained 3.35 billion parameter model optimized for efficient, strong, and balanced multilingual representation across 70+ languages, including many lower-resourced ones. The model is designed to support downstream adaptation, instruction tuning, and local deployment under realistic compute constraints.

The family ships four instruction-tuned variants, each targeting a different geographic slice:

  • tiny-aya-global: best overall balance across all 70+ languages