The context windows of current mainstream open-source models range from 32K to 128K tokens, and under long contexts the model's recall of mid-section information declines, a phenomenon the industry calls lost-in-the-middle.