Curiosity · AI Model

Aya Expanse 32B

Aya Expanse 32B, released in late 2024, is Cohere For AI's upgraded multilingual open-weights LLM. It uses an improved data synthesis and preference-training pipeline (data arbitrage and multilingual preference optimisation) to reach a new bar for open-weights multilingual quality, reportedly beating Llama 3.1 70B on many non-English evaluations at roughly half the size.

Model specs

Vendor
Cohere For AI
Family
Aya
Released
2024-10
Context window
128,000 tokens
Modalities
text

Strengths

  • State-of-the-art open-weights multilingual scores at launch
  • 128K-context window for long multilingual documents
  • Open weights under CC-BY-NC license for research

Limitations

  • Non-commercial license limits production use
  • Covers 23 languages — some major languages still out of scope
  • No vision or audio modalities

Use cases

  • Multilingual research and benchmarking
  • Region-specific assistants for underserved languages
  • Fine-tuning baselines for multilingual domain adaptation
  • Education on multilingual LLM training methods

Benchmarks

BenchmarkScoreAs of
Multilingual Arena ELO (self-reported)above Llama 3.1 70B on non-English evals2024-10
MMLU (multilingual)leads the 23 supported languages2024-10

Frequently asked questions

What is Aya Expanse 32B?

Aya Expanse 32B is Cohere For AI's 32-billion-parameter open-weights multilingual LLM, released in late 2024, trained with multilingual preference optimisation and data arbitrage for strong per-language quality.

How does Aya Expanse compare to Aya 23?

Aya Expanse improves on Aya 23 on most multilingual benchmarks thanks to stronger data curation and preference training, and ships in both 8B and 32B sizes.

Can I use Aya Expanse commercially?

The open-weights release is CC-BY-NC 4.0. Commercial deployments require a licensing agreement with Cohere; the Command R family is Cohere's commercial counterpart.

Sources

  1. Cohere For AI — Aya Expanse blog — accessed 2026-04-20
  2. Hugging Face — CohereForAI/aya-expanse-32b — accessed 2026-04-20