“Open-source models have reached a quality threshold where they are now much easier to fine-tune and steer, especially with small amounts of unique data.”