The large C model uses 3.6 billion parameters which is 6.7 GiB if each parameter is 16 bits.
The large C model have fair bit of parameters tied to text-conditioning, not to the main denoising process. Similar to how we split the network for SDXL Base, I am pretty confident we can split non-trivial amount of parameters to text-conditioning hence during denoising process, loading less than 3.6B parameters.
What's more, they can presumably be swapped in and out like the SDXL base + refiner, right?