Curiosity · AI Model

Aya 23 35B

Aya 23 35B, released by Cohere For AI in May 2024, is a multilingual open-weights LLM built on the Command R base and fine-tuned on instruction data spanning 23 languages. It was one of the strongest open-weights multilingual models at launch, outperforming Aya 101 on every covered language despite supporting fewer total languages.

Model specs

Vendor
Cohere For AI
Family
Aya
Released
2024-05
Context window
8,192 tokens
Modalities
text

Strengths

  • Strong per-language quality across 23 languages
  • Open weights under CC-BY-NC 4.0 for research use
  • Backed by Cohere For AI's open-science programme

Limitations

  • Non-commercial research license limits production use
  • Covers 23 languages — narrower than Aya 101's breadth
  • No vision or tool-use capabilities

Use cases

  • Multilingual research and academic experiments
  • Region-specific chatbots and content moderation
  • Fine-tuning baselines for underrepresented languages
  • Benchmarks for multilingual open-weights models

Benchmarks

BenchmarkScoreAs of
AyaEval multilingualstate-of-the-art open-weights at launch2024-05
MMLU (en)≈67%2024-05

Frequently asked questions

What is Aya 23 35B?

Aya 23 35B is Cohere For AI's 35-billion-parameter open-weights multilingual model, released in May 2024 with high-quality instruction tuning across 23 languages.

What is the license for Aya 23?

Aya 23 is distributed under CC-BY-NC 4.0 for research use. Commercial deployments require a separate agreement with Cohere.

How is Aya 23 different from Aya 101?

Aya 101 covered 101 languages with lower per-language quality. Aya 23 narrows to 23 languages but delivers substantially stronger per-language performance.

Sources

  1. Cohere For AI — Aya 23 announcement — accessed 2026-04-20
  2. Hugging Face — CohereForAI/aya-23-35B — accessed 2026-04-20