> For the complete documentation index, see [llms.txt](https://brindha.gitbook.io/mylearning/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://brindha.gitbook.io/mylearning/tools/aya.md).

# Aya

#### Aya: Cohere Labs' Multilingual AI Initiative

**What is Aya?**

Aya is a global open-science initiative from Cohere Labs — the nonprofit research arm of Cohere — dedicated to advancing multilingual AI and making it accessible to the communities most underserved by existing English-dominant models. Since its launch in February 2024, the Aya initiative has engaged over 3,000 independent researchers across 119 countries, resulting in the largest multilingual dataset collection to date with 513 million examples.

The core philosophy of Aya is that equitable performance across high- and low-resource languages requires diverse data, adaptive architectures, and instruction-tuned pipelines designed specifically for cross-lingual task generalization — not simply translations of English-centric models.

***

**Evolution of the Aya Model Family**

**Aya 101 (February 2024)** The first release, built on the mT5 architecture and covering 101 languages with a focus on breadth. It was the first large-scale open-source model to prioritize underserved languages through a global collaborative effort.

**Aya 23 (May 2024)** Shifted strategy from breadth to depth. Available in 8B and 35B parameter variants, built on Cohere's Command model series. By allocating more capacity to 23 well-supported languages rather than 101, it outperformed Aya 101 on those languages as well as Gemma, Mistral, and Mixtral across discriminative and generative tasks.

**Aya Expanse (October 2024)** A high-performance family in 8B and 32B sizes, using three key research innovations: synthetic data for low-resource languages, human feedback in late-stage training, and model merging. Aya Expanse 8B achieved a 60.4% simulated win rate against Google's Gemma 2 9B in m-ArenaHard multilingual benchmarks. It reduces infrastructure costs by up to 30% compared to prior approaches.

**Aya Vision (March 2025)** Cohere Labs' first multimodal model. Available in 8B and 32B sizes, supporting image captioning, visual question answering, translation, and summarization across 23 languages. The 32B model outperforms models twice its size including Meta's Llama-3.2 90B Vision on certain visual benchmarks. The 8B model outperforms models ten times its size on some evaluations. Performance gains came from synthetic annotations, multilingual data scaling, and multimodal model merging — reducing computational overhead by up to 40%. Cohere also released the AyaVisionBenchmark, a new multilingual vision evaluation suite, alongside the model. Open weights are available on Hugging Face and Kaggle.

**Tiny Aya (February 2026 — Latest)** The most recent release, launched at the India AI Summit. A 3.35B parameter model family supporting over 70 languages, capable of running fully offline on everyday devices like laptops and smartphones — achieving up to 32 tokens per second on an iPhone 17 Pro. Trained on a single cluster of 64 H100 GPUs using relatively modest compute.

The Tiny Aya family includes four specialized variants: TinyAya-Global (broad multilingual coverage), TinyAya-Earth (African and West Asian languages), TinyAya-Fire (South Asian languages including Hindi, Bengali, Tamil, Telugu, Urdu, Punjabi, Gujarati, and Marathi), and TinyAya-Water (Asia Pacific, West Asia, and European languages). A culturally aware tokenizer reduces token fragmentation across linguistic structures, improving inference efficiency.

***

**Key Technical Features**

* Context window of 8K tokens with 8K output length (Tiny Aya)
* Apache 2.0 license for language models; CC-BY-NC for Tiny Aya base (commercial use requires contacting Cohere)
* Available on Hugging Face, Kaggle, and Ollama for local deployment
* Instruction-tuned variants alongside base models for each generation
* Tokenizer designed to minimize token breakup across diverse scripts and linguistic structures

***

**Strengths**

* The most language-inclusive open model family available, covering 70 to 101 languages
* Genuine focus on low-resource and underserved language communities
* Community-driven development with 3,000+ researchers across 119 countries
* Tiny Aya enables fully offline, on-device multilingual AI without cloud dependency
* Multimodal capabilities in Aya Vision across 23 languages
* Strong benchmark performance relative to model size across multilingual tasks

**Limitations**

* Not optimized for complex reasoning, math, or code tasks
* Tiny Aya's chain-of-thought reasoning on tasks like multilingual math is comparatively weaker
* Commercial use of Tiny Aya requires contacting Cohere rather than unrestricted Apache 2.0 access
* Vision and language capabilities less deep than frontier models from larger labs
* Primarily a research-focused initiative rather than a production enterprise product