Fable is indeed larger than Sol. OpenAI is developing Astra which will be more of a Fable-sized model.

If you can train a larger model then you can distill smaller models from it. You don't need to necessarily serve the larger model publicly. Distillation is much more effective when you have unrestricted access to the original model.