News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Aya Expanse

Aya Expanse

Open-weight multilingual language models in 8B and 32B sizes from Cohere For AI

Aya Expanse is a family of multilingual language models with 8 billion and 32 billion parameters, released as open weights by Cohere For AI and Cohere. A technical report posted in December 2024 describes them as matching or beating leading open-weight models across 23 languages.

Share on WhatsApp

  • Two open-weight multilingual models: 8B and 32B parameters
  • Reported win-rate up to 76.6% against open-weight models in their class
  • 32B model reported to beat Llama 3.1 70B at a 54.0% win-rate
  • Evaluated on Arena-Hard-Auto translated into 23 languages
  • Released with the m-ArenaHard multilingual evaluation dataset

What it is

Aya Expanse is a set of two multilingual language models, one with 8 billion parameters and one with 32 billion. Cohere For AI and Cohere built them, drawing on several years of research work at both organisations. The team published a short technical report on 5 December 2024 and put the weights out openly. The goal stated in the report is to close the gap between multilingual models and models trained mainly for one language, so that speakers of many languages get similar quality to English users.

How it came about

The models come out of research at Cohere For AI, the open research arm of Cohere, and from Cohere itself. The report lists 45 authors, led by John Dang and with Ahmet Üstün as the contact author. It appeared on arXiv on 5 December 2024. Alongside the models, the team released m-ArenaHard, a multilingual evaluation dataset built by translating Arena-Hard-Auto into 23 languages.

Inside the model

The report describes three research threads feeding into the models: data arbitrage, multilingual preference training, and model merging. Data arbitrage means choosing between candidate training data sources to get better results for the same effort. Preference training tunes the model on human or synthetic judgements about which answer is better, here across many languages. Model merging blends several trained models into one. The report does not give layer counts, context length, training tokens or compute figures.

What it can do

On Arena-Hard-Auto translated into 23 languages, the report says both sizes beat leading open-weight models in their parameter class, including Gemma 2, Qwen 2.5 and Llama 3.1, with win-rates up to 76.6%. It also reports that the 32B model beats Llama 3.1 70B, a model with twice the parameters, at a 54.0% win-rate. These are the authors' own evaluation results.

How to use it

The weights are released openly, and the report is available as a preprint. The accompanying m-ArenaHard evaluation dataset was released at the same time.

Why it matters

The report presents the models as a new reference point for multilingual performance among open-weight models. The released m-ArenaHard dataset gives other researchers a way to test multilingual ability across 23 languages.

Questions people ask about Aya Expanse

What is Aya Expanse?

A pair of multilingual language models with 8 billion and 32 billion parameters, released as open weights by Cohere For AI and Cohere. A technical report describing them was posted in December 2024, along with a multilingual evaluation dataset called m-ArenaHard.

How many languages does it support?

The report evaluates the models on Arena-Hard-Auto translated into 23 languages. It does not list the languages by name or give a separate supported-language count.

How does it compare to other open models?

The authors report that both sizes beat Gemma 2, Qwen 2.5 and Llama 3.1 in their parameter classes on the translated Arena-Hard-Auto set, with win-rates up to 76.6%. They also report the 32B model beating Llama 3.1 70B at 54.0%.

Can I download it?

Yes. The report states that the weights are released openly, alongside the m-ArenaHard evaluation dataset.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Aya Expanse yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2412.04261. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.