Llama3 70B Instruct

Llama 3 70B Instruct is the larger instruction-tuned model in Meta's Llama 3 release, optimized for dialogue and conversational tasks. Trained on over 15 trillion tokens from publicly available sources with a December 2023 knowledge cutoff, it combines supervised fine-tuning with reinforcement learning from human feedback to align with user preferences. The model is strong at reasoning, code generation, and instruction-following, with reported benchmark scores of 82.0 on MMLU and 81.7 on HumanEval. It supports an 8,192-token context window and uses Grouped-Query Attention for more efficient inference. The weights are openly available for both commercial and research use, and the model went through extensive red-teaming for safety. Compared to its 8B sibling, the 70B model offers substantially higher reasoning and coding capability in exchange for greater computational requirements, making it suited to more complex tasks where the added quality justifies the resources.