Fewer parameters at inference time makes a massive difference in cost for batch jobs, assuming vram usage is the same