Cohere Aya 3 is a strong contender for multilingual tasks and benefits from its open-source nature and community support. GPT-6, with its anticipated massive scale and focus on ethical guidelines, promises advanced capabilities but lacks specific details on community support and has a different pricing model.
| Attribute | Cohere Aya 3 | GPT-6 |
|---|---|---|
| Context Window Length | 8K (8192) tokens for Aya 23, 16K for Aya Vision, 128k for Aya Expanse 32B | Up to 1 million tokens plausible in the near future. |
| Number of Parameters | 8 billion and 32 billion parameters for Aya Expanse, 8 billion and 35 billion parameter versions for Aya 23 | Estimates suggest training models up to 80 trillion parameters. Hardware exists to train a 27T parameter model, and 50T parameters isn't out of the question. Microsoft anticipates needing 'two orders of magnitude more computation' than GPT-5. |
| Training Data Size | The Aya Collection has 513 million prompts and completions across 114 languages. | Quadrillions of tokens. |
| Multilingual Support (Number of Languages) | Aya 23 supports 23 languages, Aya 101 covered 101 languages. Languages include Arabic, Chinese (simplified & traditional), Czech, Dutch, English, French, German, Greek, Hebrew, Hindi, Indonesian, Italian, Japanese, Korean, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Turkish, Ukrainian, and Vietnamese. | ChatGPT already supports over 95 languages; improved cross-lingual abilities, including low-resource languages, are expected. |
| Finetuning Capabilities | Aya models are instruction fine-tuned (IFT) to follow human instructions. | Extending the finetuning API to the latest models is a priority. Fine-tuning GPT models for specific domains enhances contextually accurate content generation. |
| Inference Speed (Tokens/Second) | Aya Expanse 32B: 122 tokens per second (median), Aya Expanse 8B: 167.7 tokens per second | NVIDIA Blackwell can deliver 30x more throughput at reading speeds of 20 tokens per user per second using specific configurations. User experience depends on response time, measured in tokens per second per user. |
| API Availability and Pricing | Aya Expanse models (8B and 32B): $0.50 per 1M tokens (input), $1.50 per 1M tokens (output). Free tier with rate-limited usage available. Production API keys available with pay-as-you-go pricing. | OpenAI aims to drive the cost of intelligence down, working to reduce API costs over time. The API plan costs $20 per 1,000 interactions, with a minimum of $50 per month, plus token consumption. |
| Code Generation Performance (e.g., HumanEval Score) | Not available | GPT-J's performance is better than GPT-3 in writing code. |
| Hallucination Rate (Percentage) | Not available | Hallucination rates can significantly impact the reliability of AI systems; they are decreasing with larger models. |
| Bias Evaluation (e.g., Gender Bias Score) | Tested for toxicity and bias in open-ended generation and gender bias in translation. Racial and gender biases are still present despite mitigation efforts. | Incorporation of ethical guidelines in the model s development process to address bias, fairness, and transparency proactively in its training and outputs is expected. |
| Safety Measures and Red Teaming Results | Multilingual safety measures, context distillation to generate refusal messages for unsafe contexts. Community-based red-teaming possible due to open-source nature. | Enhanced safety features are expected to minimize misuse risks, including better detection of harmful content and misinformation. Red Teaming is used to identify vulnerabilities. |
| Community Support and Documentation Quality | Community support through Discord and documentation available. | Not available |