In this video, we do a deep dive on Cohere's latest multilingual models, a family of models called Tiny Aya, which specialize in multilingual uses at the edge. 

Blog: https://cohere.com/blog/cohere-labs-tiny-aya
HF: https://huggingface.co/collections/CohereLabs/tiny-aya
Colab:  https://dripl.ink/rV4gP

For more tutorials on using LLMs and building agents, check out my Patreon
Patreon: https://www.patreon.com/SamWitteveen
Twitter: https://x.com/Sam_Witteveen 

🕵️ Interested in building LLM Agents? Fill out the form below
Building LLM Agents Form: https://drp.li/dIMes

👨‍💻Github:
https://github.com/samwit/llm-tutorials

⏱️Time Stamps:
00:00 Intro
01:30 LlaMA 2 Tokenizers
03:51 TinyAya Blog
04:47 TinyAya Models
07:42 TinyAya Paper
09:08 TinyAya in HuggingFace
09:42 TinyAya Colab Demo

Sam Witteveen AI is a publication offering insights, tutorials, and resources for artificial intelligence (AI) enthusiasts and practitioners. Readers can learn about machine learning algorithms, deep learning frameworks, and AI applications. With tutorials, case studies, and expert interviews, Sam Witteveen AI provides  guidance and expertise for building and deploying AI solutions.

Sam Witteveen

Cohere has released the Aya Tiny model family — a suite of small (~3.3B parameter) multilingual language models designed to support 70+ languages, including many low-resource ones underserved by mainstream models. The release includes a base model, a general-purpose 'Global' model, and three region-specialized variants (Earth for West Asia/Africa/Europe, Fire for South Asia, Water for Asia-Pacific). Each specialized model is built by merging region-specific SFT checkpoints. Cohere trained a custom tokenizer optimized for multilingual efficiency, competitive with Gemma 3 and Qwen 3 tokenizers. The models are small enough to run on mobile devices, making them practical for building assistants in languages not covered by larger models. Multilingual training datasets and benchmarks are also released to support fine-tuning.

Tiny Aya - Cohere's Mini Multilingual Models