Sharding
Splitting one large model file into numbered parts. You need all of them.
Files above a few gigabytes are commonly split into parts named like model-00001-of-00009. Every part is required — the first is not a smaller version of the model, and a runtime given only one will fail to load rather than run a truncated model.
When we report a quantization's size we sum every shard. Recording one and calling it the model's size is off by the shard count, which for some repositories is thirty.