Aya Expanse 8B

Overview

Aya Expanse 8B is the 8-billion-parameter model from the Aya Expanse release by Cohere Labs (formerly Cohere For AI), which also includes a 32B version. Cohere paired a pre-trained model from its Command family with a year of multilingual research that the card lists as data arbitrage, multilingual preference training, safety tuning and model merging. It reads and writes text only, with an 8K-token context window.

The release post describes the arbitrage step: a pool of teacher models, each trained for a group of languages, answers every prompt, and a reward model keeps the best completion, since no single teacher is strong in all languages. Models trained on different language families were then merged by weighted averaging.

The 23 supported languages are Arabic, Chinese (simplified and traditional), Czech, Dutch, English, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Turkish, Ukrainian and Vietnamese. The card compares it with Gemma 2 9B, Llama 3.1 8B, Ministral 8B and Qwen 2.5 7B on win rates judged by GPT-4o, including on m-ArenaHard, a translation of Arena-Hard-Auto into all 23 languages.

What you can build

• Chat assistants for research or non-commercial projects that answer each user in their own language.
• Letters and personal messages written directly in the target language. The card's usage example is a Turkish prompt asking for a letter to one's mother.
• Questions answered from a short document pasted into the prompt, within the 8K-token window.

Inputs

The prompt is the only required field. An optional system prompt, filled in by default, goes before it to set role and tone.

For multi-turn chat, add a user ID and a session ID. Wiro stores the history per user, and each new session ID opens a separate thread for that user. Both fields can stay empty.

You can also set temperature (0 to 5, default 0.7), top-p, top-k, repetition and length penalties, token limits, semicolon-separated stop sequences and a seed. Two switches start on: sampling, which adds randomness when the model picks each token, and quantization, which loads the weights at reduced precision to save memory and may slightly lower output quality.

Outputs

Plain text. The default prompt asks about historical events near the Tower of London and how they could inspire fiction. The sample reply lists eight, starting with Anne Boleyn's execution in 1536, each followed by a story idea.

Recommended settings

Cohere's usage example samples at temperature 0.3, well below the 0.7 default here. For translation and factual answers, go near 0.3 or switch sampling off. Keep the default for brainstorming and fiction.

Limitations

• A long chat thread or a pasted document can fill the 8K window. Start a new session ID or shorten the input.
• The card makes no claims for languages beyond the 23 listed.
• It can state wrong facts with confidence. The sample here lists the 1851 Great Exhibition among events near the Tower, although it was held in Hyde Park, several kilometers west. Check names and dates before publishing.

Safety & compliance

The weights are released under CC-BY-NC, a non-commercial license, and use also requires following Cohere Labs' Acceptable Use Policy. Paid products and other commercial uses fall outside these terms.