Very useful, especially for static pre-2022 information. If these models were <5GB in size, I'd happily have the model always sitting in memory on my raspberry Pi waiting for me to ask it something.

For newer data though, you cannot beat the army of trawlers and the constant retraining that the commercial models offer... though given the nature of the steadily poisoned information sources post-2022, their effectiveness in this is steadily diminishing.