Tiny Aya

Tiny Aya is a compact 3.35-billion parameter multilingual model that achieves state-of-the-art translation quality, strong multilingual understanding, and high-quality target-language generation across 70 languages. Despite its small size, Tiny Aya is designed to run locally on any device — including phones — without requiring cloud dependency.

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.

Tiny Aya supports 70 languages: English, Dutch, French, Italian, Portuguese, Romanian, Spanish, Czech, Polish, Ukrainian, Russian, Greek, German, Danish, Swedish, Norwegian, Catalan, Galician, Welsh, Irish, Basque, Croatian, Latvian, Lithuanian, Slovak, Slovenian, Estonian, Finnish, Hungarian, Serbian, Bulgarian, Arabic, Persian, Urdu, Turkish, Maltese, Hebrew, Hindi, Marathi, Bengali, Gujarati, Punjabi, Tamil, Telugu, Nepali, Tagalog, Malay, Indonesian, Vietnamese, Javanese, Khmer, Thai, Lao, Chinese, Burmese, Japanese, Korean, Amharic, Hausa, Igbo, Malagasy, Shona, Swahili, Wolof, Xhosa, Yoruba, and Zulu.

Tiny Aya is released as a family of five models: a pretrained foundation model, a globally balanced instruction-tuned variant, and three region-specialized variants.

Tiny Aya excels at the following tasks:

• Translation: State-of-the-art multilingual translation quality across 70 languages.
• Multilingual understanding: Strong performance on multilingual comprehension benchmarks.
• Target-language generation: High-quality text generation in the target language.
• On-device deployment: Compact enough to run locally on phones and edge devices.

The instruction-tuned Tiny Aya variants (tiny-aya-global, tiny-aya-earth, tiny-aya-fire, tiny-aya-water) are available on the Cohere API via the Chat endpoint. You can use them with the Cohere SDK just like any other model:

import cohere

co = cohere.ClientV2("<YOUR_API_KEY>")

response = co.chat(
    model="tiny-aya-global",
    messages=[
        {
            "role": "user",
            "content": "Bonjour! Pouvez-vous me raconter une courte histoire en français?",
        }
    ],
)

print(response.message.content[0].text)

Tiny Aya is also available as open-weight models through Hugging Face. You can download and run the models locally:

• tiny-aya-base
• tiny-aya-global
• tiny-aya-earth
• tiny-aya-fire
• tiny-aya-water

GGUF quantized versions are also available for efficient local inference:

• tiny-aya-global-GGUF
• tiny-aya-earth-GGUF
• tiny-aya-fire-GGUF
• tiny-aya-water-GGUF

• Tiny Aya research paper
• Blog post: Cohere Labs Launches Tiny Aya
• Aya research initiative