Llama 3.2 includes a 1B parameter model. This should be 8x higher throughput for data pipelines. In our experience, smaller models are just fine for simple tasks like reading paragraphs from PDF documents.