Aya

Aya is an open-science initiative from Cohere Labs that produces a family of open-weights multilingual models spanning text-only and multimodal tasks.

Reviewed by 7wData

On this page

Publisher review

Aya is an open-science initiative from Cohere Labs that produces a family of open-weights multilingual models spanning text-only and multimodal tasks. It targets researchers and developers who need strong language coverage across 101 languages without relying on proprietary APIs. The project includes three distinct research models: Aya Vision (8B and 32B parameter multimodal models), Aya Expanse (a text-only model mastering 101 languages through instruction tuning), and Aya 101 (an earlier instruction-tuned model covering 101 languages). These models are designed for translation, cross-lingual transfer, and vision-language tasks, with particular emphasis on languages underrepresented in existing LLM training data.

Aya Vision achieves leading performance across both multimodal and text-only tasks, outperforming significantly larger models on benchmarks spanning 23 languages. It uses a synthetic annotation framework to curate high-quality multilingual multimodal instruction data, and a cross-modal model merging technique that mitigates catastrophic forgetting — preserving text-only capabilities while adding vision. Aya Expanse applies innovative instruction tuning and cross-lingual transfer techniques to handle 101 languages, reducing computational overhead by up to 40% via optimized training. The models are released under open weights, allowing fine-tuning and deployment on local infrastructure.

In the open-weight multilingual space, Aya competes with Mistral and Llama 3, both of which offer strong English-centric performance but narrower language coverage. Mistral provides 7B and mixture-of-experts models (8x7B Mixtral, 8x22B Mixtral) under Apache 2.0, while Llama 3 offers 8B and 70B parameter models under a custom commercial license. Aya differentiates by explicitly targeting 101 languages and providing multimodal capabilities in the Vision variant, though it lacks the extensive ecosystem and community tooling that both Mistral and Llama 3 have accumulated.

The honest trade-off: Aya's broad language coverage comes at the cost of English-centric benchmark performance — it is not designed for coding or general-purpose tasks. The models are research-focused, meaning they may require more manual tuning for production use. The open-weights release enables customization but also means users must manage their own infrastructure. Additionally, the project's documentation and community support are less mature than those of Mistral or Llama 3, which have larger user bases and more third-party integrations.

Get the AI & data signal, daily.

335k+ subscribers read this every morning. One email, both newsletters. Unsubscribe anytime.

How it works

  1. Multilingual coverage

    Supports 101 languages, including many underrepresented ones, through instruction tuning and cross-lingual transfer techniques.

  2. Multimodal capability

    Aya Vision handles both images and text, achieving state-of-the-art performance across 23 languages on multimodal benchmarks.

  3. Open-weights release

    Models are available under open weights, enabling local deployment, fine-tuning, and research without vendor lock-in.

  4. Synthetic data generation

    Uses a synthetic annotation framework to create high-quality multilingual multimodal instruction data, addressing data scarcity.

  5. Cross-modal model merging

    Merges vision and text models to mitigate catastrophic forgetting, preserving text-only performance while adding vision capabilities.

  6. Optimized training efficiency

    Reduces computational overhead by up to 40% through optimized training techniques compared to baseline multilingual models.

  7. Research-first design

    Part of Cohere's open-science initiative, with detailed technical reports and reusable components for the research community.

Strengths and trade-offs

Strengths

  • Achieves leading performance across both multimodal and text-only tasks, outperforming significantly larger models on 23-language benchmarks.
  • Covers 101 languages through instruction tuning, far exceeding the typical 10-20 language range of most open-weight models.
  • Reduces computational overhead by up to 40% via optimized training techniques, making it more accessible for resource-constrained teams.
  • Uses synthetic data generation and cross-modal merging to address data scarcity and catastrophic forgetting in multilingual settings.

Trade-offs

  • Not designed for coding or general-purpose tasks; primary focus is translation and cross-lingual transfer.
  • Research-focused models may require additional tuning and infrastructure management for production deployment.
  • Community support and documentation are less mature compared to Mistral or Llama 3, which have larger user ecosystems.
  • English-centric benchmark performance lags behind specialized English-only models of similar parameter counts.

Pricing context

No pricing information available; models are open-weights and free to use under the provided license.

Getting started with Aya

  1. Download model weights

    Visit the official Aya repository on Hugging Face or GitHub. Select the desired model variant (Aya Vision, Aya Expanse, or Aya 101) and download the open-weights files to your local machine or server.

  2. Set up Python environment

    Create a Python virtual environment with Python 3.10 or later. Install the required dependencies including PyTorch, Transformers, and Accelerate libraries using pip. Ensure your environment has sufficient GPU memory for the model size.

  3. Load the model

    Use the Transformers library to load the downloaded model and tokenizer. For example, call `AutoModelForCausalLM.from_pretrained()` with the local path to the model directory. Set the device map to 'auto' for multi-GPU support.

  4. Run a translation task

    Prepare a text input in one of the 101 supported languages. Tokenize the input using the loaded tokenizer, generate output with `model.generate()`, and decode the response. Verify the translation quality for your target language pair.

  5. Fine-tune for production

    Collect a domain-specific multilingual dataset and format it for instruction tuning. Use the Hugging Face Trainer API or a custom training loop to fine-tune the model on your data. Save the fine-tuned weights and deploy using a serving framework like vLLM.

Frequently Asked Questions

What is Aya and who created it?

Aya is an open-science initiative from Cohere Labs that produces a family of open-weights multilingual models. It includes text-only and multimodal variants covering 101 languages, targeting researchers and developers who need strong language coverage without relying on proprietary APIs.

How many languages does Aya support?

Aya supports 101 languages through instruction tuning and cross-lingual transfer techniques. This includes many underrepresented languages that are often missing from other large language models, making it a strong choice for multilingual tasks like translation and cross-lingual transfer.

What are the different Aya models available?

Aya includes three research models: Aya Vision (8B and 32B parameter multimodal models), Aya Expanse (a text-only model mastering 101 languages), and Aya 101 (an earlier instruction-tuned model). Each serves different needs, from vision-language tasks to pure text translation.

How does Aya compare to Mistral and Llama 3?

Aya offers broader language coverage with 101 languages, while Mistral and Llama 3 focus on English-centric performance with narrower language support. However, Aya lags in English benchmark performance and has less mature community support and documentation compared to its competitors.

Can I use Aya for free and deploy it locally?

Yes, Aya models are released under open weights, allowing free use, fine-tuning, and local deployment without vendor lock-in. There is no pricing information available, and users must manage their own infrastructure for production use.

What are the main weaknesses of Aya models?

Aya is not designed for coding or general-purpose tasks, and its English-centric benchmark performance lags behind specialized English-only models. The research-focused models may require additional tuning for production, and community support is less mature than Mistral or Llama 3.

Alternatives

How Aya compares

Direct head-to-head against 2 competitors. Picked by 7wData.

This tool

Aya

Pricing
No pricing information available; models are open-weights and free to use under the provided license.
Target
Aya is an open-science initiative from Cohere Labs that produces a family of open-weights multilingual models spanning text-only and multimodal tasks.
Strength
Achieves leading performance across both multimodal and text-only tasks, outperforming significantly larger models on 23-language benchmarks.
Watch for
Not designed for coding or general-purpose tasks; primary focus is translation and cross-lingual transfer.

Command R

Pricing
Custom/Contact sales
Target
Enterprises needing accurate, citation-backed text generation
Deployment
Cloud API
Strength
Clear citations to reduce hallucinations
Watch for
Pricing not publicly disclosed; may escalate with usage

Falcon 2

Pricing
Free (open-source)
Target
Developers needing open-source multilingual multimodal models
Deployment
Self-hosted
Strength
Vision-to-language capabilities in an 11B model
Watch for
Smaller community and fewer integrations than Llama 3

User reviews

No user reviews yet. Be the first to write one.

Sources

Reporting on this tool draws on these publicly available sources.

  1. www.reddit.com
  2. openreview.net
  3. cohere.com
  4. sapling.ai
  5. blog.promptlayer.com