“Open-source models are often slow out of the box. Model providers then optimize these models to make them fast and release them behind proprietary APIs, without contributing the improvements back to the community.”
Replicate gets FLUX.1 [schnell] to 0.29 seconds at 512x512 by quantizing to fp8, which it admits “slightly changes the output of the model.” The shortcut is on by default and you opt out by setting go_fast to false. Publishing the code costs little when the business is renting out the GPUs that run it.