Nvidia has quietly released an open-source AI language model, Nemotron 70B, which is a fine-tuned version of Llama 3.1 70B. According to benchmarks, Nemotron 70B outperforms OpenAI's GPT-4o and Anthropic's Claude 3.5 Sonnet on several tests, including Arena Hard, AlpacaEval 2 LC, and MT-Bench. On the Arena Hard benchmark, Nemotron scored 85.0, surpassing GPT-4o's 79.3 and Claude 3.5 Sonnet's 79.2. In the AlpacaEval 2 LC test, it achieved a score of 57.6, compared to GPT-4o's 57.5 and Claude 3.5 Sonnet's 52.4. On MT-Bench, Nemotron scored 8.98, outperforming GPT-4o's 8.74 and Claude 3.5 Sonnet's 8.81. The model was trained using Reinforcement Learning from Human Feedback (RLHF) with the REINFORCE algorithm and HelpSteer2-Preference prompts. Nvidia has made the instruct model, reward model, and training data available for free on Hugging Face.