Aya 23 8B

Aya 23 8B, published on 23 May 2024, is the pivot point of Cohere Labs' multilingual programme. Its predecessor Aya 101 covered 101 languages with a single 12.9-billion-parameter model and was, by the lab's own account, spread too thin; Aya 23 kept 23 widely spoken languages - together the first language of roughly half the world's population, Polish among them - and gave each of them a far larger share of the model's capacity. The trade is visible in the numbers alone: a fifth of the language count in a smaller model. The 8B variant holds 8,028,033,024 parameters and sits on Cohere's Command R architecture rather than the mT5 encoder-decoder that Aya 101 had been fine-tuned from - a change of family, not just of scale. Its parameter count is identical to Aya Expanse 8B, released five months later, which is the clearest evidence that the later model is a retrained version of the same backbone rather than new hardware for the task. The licence is CC BY-NC 4.0 with the acceptable use policy, weights are gated behind an account and acceptance of the terms, and Cohere does not host the model. It remains in use two years on, with 10,516 downloads in the thirty days to 18 August 2026 and 436 likes - second only to Aya Expanse 8B across the family.

▸Videos

No videos yet.

▸Reviews