Aya 101

Aya 101 is the first model of Cohere Labs' Aya programme, published on 8 February 2024 alongside the paper "Aya Model: An Instruction Finetuned Open-Access Multilingual Language Model". Its ambition was coverage rather than benchmark scores: it follows instructions in 101 languages, roughly twice as many as the multilingual models it was measured against at the time, mT0 and BLOOMZ, and it covers a long tail of low-resource languages that commercial models still skip. Polish is on the list, classified by the authors as a high-resource language. Architecturally it is not one of Cohere's own trunks. Aya 101 is an instruction fine-tune of Google's mT5-XXL - an encoder-decoder T5 model rather than the decoder-only design that became standard afterwards - carrying 12.92 billion parameters in the published checkpoint. The fine-tuning is documented unusually openly: 25 million samples seen, batch size 256, trained on a TPUv4-128 pod with T5X and Jax. The training mixture combines the xP3x collection, the crowdsourced Aya Dataset and Aya Collection, a commercially-verified subset of the Data Provenance collection, and ShareGPT-Command; because the backbone is mT5, the language list is exactly the 101 languages mT5 supports. Being a T5, it uses relative position encoding and publishes no fixed context window. What makes Aya 101 worth a separate profile two years on is its licence. Cohere Labs released the checkpoints under Apache 2.0 - the company states this explicitly as a mission choice - while every later Aya release, from Aya 23 through Aya Expanse and Aya Vision to the Tiny Aya family of 2026, carries CC BY-NC 4.0 and, in the newer cases, an access request in front of the weights. Aya 101 remains downloadable without a gate. The community response tracks that: it is still the most-liked repository in the whole family, ahead of the newer and technically stronger Aya Expanse models, despite being the oldest. By 2026 standards the model is dated - a 2024 encoder-decoder that is outperformed by far smaller modern multilingual models - but it is the reference point every later Aya release is measured against, and the only one a commercial project may legally build on.

▸Videos

No videos yet.

▸Reviews