Human evaluators also rated Llama 3 higher than other models, including OpenAI’s GPT-3.5. Meta created a new dataset for human evaluators to emulate real-world scenarios where Llama 3 might be used, including advice, summarization, and creative writing. The company is currently training larger versions of Llama 3, with over 400B parameters, which are expected to understand longer strings of instructions and data and be capable of more multimodal responses. However, Meta has not yet released a preview of these larger models.
Key takeaways:
- Meta's next generation large language model, Llama 3, has been released to cloud providers and model libraries, offering improved performance over most current AI models.
- Llama 3 features two model weights, with 8B and 70B parameters, and has shown more diversity in answering prompts, fewer false refusals, and better reasoning capabilities.
- In benchmark testing, both sizes of Llama 3 outperformed similarly sized models like Google’s Gemma and Gemini, Mistral 7B, and Anthropic’s Claude 3.
- Meta is currently training larger versions of Llama 3, with over 400B parameters, which are expected to understand longer strings of instructions and data and be capable of more multimodal responses.