Published evaluations or benchmarks from AI research papers or Meta

Meta Launches Llama 3.3 70B Model Matching GPT-4 Performance at a Fraction of the Cost

Dec 6, 2024, 06:03 PM

On December 6, 2024, Meta released Llama 3.3, a new AI language model offering state-of-the-art performance at a significantly reduced cost. The 70B-parameter model achieves performance comparable to GPT-4, with an MMLU score of 86, and supports a context length of 128K tokens. Llama 3.3 demonstrates a cost reduction factor as high as 450 times over the past year, with input token costs as low as $0.10 per million tokens, substantially undercutting competitors like GPT-4o ($2.5 per million tokens), Claude 3.5 ($3.0 per million tokens), and Gemini Pro ($1.3 per million tokens). The model is already being integrated into platforms such as HuggingChat, GroqCloud, Hyperbolic, and others, making high-quality AI more accessible and cost-effective. The release signifies a major advancement in making state-of-the-art language models widely available.