December 6, 2024:
Meta Unveils Cost-Efficient Llama 3.3 70B Model - Meta Platforms Inc. has launched Llama 3.3 70B, an open-source large language model that matches the quality of Llama 3.1 405B but requires less hardware. By optimizing the Transformer architecture and using Nvidia H100 chips, Llama 3.3 70B reduces infrastructure costs by nearly five times. It was trained on 15 trillion tokens and refined with supervised fine-tuning and RLHF techniques.
In benchmarking, Llama 3.3 70B nearly matches Llama 3.1's performance and surpasses OpenAI's GPT-4o, all at a significantly lower cost. The model is accessible on Hugging Face.
