The open-source AI landscape has a new, serious contender for multilingual tasks. Cohere's release of the Aya 23 family, with 8B and 35B parameter open-weight models, provides a much-needed, high-performance baseline for builders working in the 23 languages it covers. This isn't just another model drop; it's a practical alternative to relying on closed APIs or fine-tuning English-centric models for global applications.
what is aya 23
Aya 23 is a family of instruction-tuned, decoder-only transformer models released by Cohere for AI, the company's non-profit research lab. It comes in two sizes: an 8-billion parameter model designed for accessibility and a larger 35-billion parameter version for more complex tasks.
This release represents a strategic shift from its predecessor, Aya 101, which aimed for breadth across 101 languages. Aya 23 instead focuses on depth, allocating more training capacity to a curated list of 23 languages, including Arabic, Chinese, German, Hindi, Japanese, Spanish, and Vietnamese. The goal is to provide state-of-the-art capabilities for a set of languages that cover roughly half the world's population.
The models are based on Cohere's Command series and were fine-tuned on the Aya Collection dataset. By releasing the model weights, Cohere allows researchers and developers to inspect and build on top of their work, a move that distinguishes it from fully closed-source offerings.
why this matters for builders
For engineers building products for non-English speaking markets, the options have often been limited. You could use a proprietary, closed-source API, which offers high performance but limited customizability and potential lock-in. Or, you could take a powerful open-source but English-centric model and attempt to fine-tune it, with performance often lagging in other languages.
Aya 23 offers a compelling middle ground. The 8B model, in particular, is designed to be accessible, running on consumer-grade hardware, which significantly lowers the barrier to entry for developers and researchers. The 35B model provides a more powerful option that benchmarks show outperforms other popular open models like Gemma and Mistral on a range of multilingual tasks.
This enables more robust applications in areas like multilingual customer support, content moderation, and language learning tools without starting from scratch. Having a strong, open baseline model that is already pre-trained on a diverse set of languages saves significant computational cost and data sourcing effort.
getting started and considerations
You can access the models on Hugging Face. The weights are available under a CC-BY-NC license, which is permissive for research but has non-commercial restrictions. Here's a quick example of how you might run inference with the 8B model using the transformers library.
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
tokenizer = AutoTokenizer.from_pretrained("CohereForAI/aya-23-8B")
model = AutoModelForCausalLM.from_pretrained("CohereForAI/aya-23-8B", torch_dtype=torch.bfloat16)
# Format the prompt using the ChatML template
# Each message is a dictionary with 'role' and 'content'
messages = [
{"role": "system", "content": "You are a helpful assistant."},
{"role": "user", "content": "Wie ist das Wetter heute in Berlin?"}
]
input_ids = tokenizer.apply_chat_template(
messages,
tokenize=True,
add_generation_prompt=True,
return_tensors="pt"
)
# Generate a response
outputs = model.generate(input_ids, max_length=256)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)
A few things to keep in mind:
- Hardware: While the 8B model is more accessible, the 35B version still requires significant computational resources. Quantization will likely be necessary for running it on local or consumer-grade hardware.
- Licensing: The Creative Commons non-commercial license means you need to consider your use case. It's ideal for research, experimentation, and internal tools, but commercial applications may require a different approach.
- Evaluation: Benchmarks are useful, but always evaluate the model's performance on your specific tasks and target languages. Performance can vary, and what works for a general benchmark may not hold for your specific domain.
the takeaway
The release of Aya 23 is a meaningful step toward democratizing high-performance AI beyond English. It provides builders with a powerful, open set of tools to create more globally relevant and linguistically inclusive applications. For teams that have been hampered by the cost of proprietary APIs or the performance limitations of English-first open models, Aya 23 is a development worth investigating.
Top comments (0)