“The performance of the 32-billion parameter Olmo model is comparable to the original version of GPT-4.”