- Runs
ollama show <model> --modelfileand finds the GGUF blob that containsclef.*tensors. - Moves the
clef.*tensors to the front of the data section, which puts them below 4 GiB, and reorders the tensor table to match. - Runs
ollama create <new-name>with the originalModelfileand the new path, then deletes the temporary copy. - If the head is already below 4 GiB, it exits without doing anything.
It needs about 10 GB of temporary disk (TEMP or TMPDIR) plus about 10 GB for the new model.