Cost-efficiency analysis reports from AI research institutions

Meta Launches Open-Source Llama 3.1 with 405B Parameters, Challenging GPT-4o

Jul 24, 2024, 05:06 PM

Meta has launched Llama 3.1, its latest open-source AI model, which is being touted as a significant advancement in the AI landscape. The Llama 3.1 family includes models with 8 billion, 70 billion, and 405 billion parameters. These models are optimized for various tasks, including synthetic data generation, distillation, translation, and coding. They feature 128K context length and multilingual support. The models are available for deployment on both NVIDIA and AMD GPUs, with NVIDIA TensorRT-LLM optimization offering superior throughput and latency. Fireworks AI is among the first to offer Llama 3.1 inference using both NVIDIA and AMD GPUs. Yann LeCun, a prominent figure in AI, claims that Meta AI is quickly becoming the most widely used assistant, challenging the dominance of OpenAI's GPT-4o. Llama 3.1 models are also noted for their cost efficiency and performance, rivaling closed-source models. Mark Zuckerberg, CEO of Meta, has emphasized the importance of open-source AI, stating that it allows for greater customizability and innovation. The release of Llama 3.1 has intensified the AI arms race, with OpenAI responding by offering free fine-tuning for its GPT-4o Mini model.