A few weeks ago, NVIDIA released Nemotron, a large language model that they derived from Meta’s Llama 3.1 70B. NVIDIA claimed at release that Nemotron outperformed GPT-4o and Claude 3.5 Sonnet on certain benchmarks. That was exciting news, as my experience with self-hostable AI models is that they trail commercial models by about a year in terms of accuracy and quality. I decided to test out Nemotron with a few simple coding tasks to see how it compared to commercial models like Claude 3.