Aya Expanse
CohereAya Expanse is a family of multilingual chat models from Cohere Labs, Cohere's research lab, released in October 2024 at 8B and 32B parameters. It covers 23 languages and was built by combining data arbitrage, multilingual preference training and model merging. Cohere's later Tiny Aya and Aya Vision models are separate lines.
Openness
2 high confidence- weights
- open (8B and 32B on Hugging Face behind an automatically approved access form)
- data
- closed
- code
- closed (no training or post-training pipeline published)
- license
- CC-BY-NC-4.0 (non-commercial, with Cohere Labs' acceptable-use addendum)
Aya Expanse's weights can be downloaded after an automatically approved form, but Cohere licenses them under CC-BY-NC 4.0 with an acceptable-use addendum, which rules out commercial use. The training data and recipe are not released.
- https://arxiv.org/abs/2412.04261 recorded 2026-11-12
Abstract: "we present extended evaluation results for the Aya Expanse model family and release their open-weights", and names the training advances ("data arbitrage, multilingual preference training, and model merging") without releasing training data or code.
- https://cohere.com/c4ai-cc-by-nc-license recorded 2026-11-12
Cohere's "CC-BY-NC 4.0 License with Acceptable Use Addendum": "Creative Commons Attribution-NonCommercial 4.0 International Public License with Acceptable Use Addendum", which the Aya Expanse access form asks users to accept.
- https://huggingface.co/api/models/CohereLabs/aya-expanse-32b recorded 2026-11-12
Model API for CohereLabs/aya-expanse-32b: `gated: "auto"`, cardData.license `cc-by-nc-4.0`.
- https://huggingface.co/api/models/CohereLabs/aya-expanse-8b recorded 2026-11-12
Model API for CohereLabs/aya-expanse-8b: `gated: "auto"`, cardData.license `cc-by-nc-4.0`.
Adoption
2 high confidenceHugging Face downloads summed across the 8B and 32B checkpoints, most of them for the 8B model.
- https://huggingface.co/api/models/CohereLabs/aya-expanse-32b recorded 2026-11-12
"downloads":3266 for CohereLabs/aya-expanse-32b, the trailing 30-day count the Hub API reports.
- https://huggingface.co/api/models/CohereLabs/aya-expanse-8b recorded 2026-11-12
"downloads":23586 for CohereLabs/aya-expanse-8b, the trailing 30-day count the Hub API reports.
Capability
3 medium confidenceCohere reports Aya Expanse 32B beating Llama 3.1 70B, and the 8B model beating Llama 3.1 8B and Gemma 2 9B, on multilingual Arena-Hard comparisons, which puts the family level with Llama's instruct models of the same generation. Those are pairwise win rates judged by another model, and no general benchmark places it near current releases.
- https://arxiv.org/abs/2412.04261 recorded 2026-11-12
Abstract: "Aya Expanse 8B and 32B outperform leading open-weight models in their respective parameter classes, including Gemma 2, Qwen 2.5, and Llama 3.1, achieving up to a 76.6% win-rate. Notably, Aya Expanse 32B outperforms Llama 3.1 70B ... achieving a 54.0% win-rate."
- https://huggingface.co/blog/aya-expanse recorded 2026-11-12
Release post: "Aya Expanse 32B outperforms Gemma 2 27B, Mistral 8x22B, and Llama 3.1 70B" and Aya Expanse 8B "outperforms ... Gemma 2 9B, Llama 3.1 8B, and the recently released Ministral 8B with win rates ranging from 60.4% to 70.6%", on Arena-Hard-Auto translated into 23 languages.
Verified 2026-11-12