Because it has fewer parameters and (in some cases) it’s quantized. The hardware needed to run local inference on the full model is not really feasible to most people. Though, the release of it will probably still make a wide impact on the quality of other upcoming smaller models being distilled from it, or trained on synthetic data from it, or merged with it, etc.

You’re asking the real important questions!

And it’s great for sorting by date.

Thanks for sharing. It seems like there’s a lot of supported options. Many of them, I have no idea what are, but cars and doorbells are easy enough to understand, at least. Do you have any examples of interesting, less obvious use cases of your own, or of others’?

For local reception, receivers with RTL2832U chips are a cheap option. They are also called RTL-SDR. I have simply been using a long wire as a “random wire antenna”. Some of the older dongles also need an upconverter to be able to tune into low HF frequencies:

An upconverter for the RTL-SDR translates low HF frequencies ‘up’ into ones that are receivable by the RTL-SDR. This is a different method to the direct sampling mode used in the V3 dongles to achieve HF reception.

Quoted source: https://www.rtl-sdr.com/a-homebrew-one-transistor-upconverter-for-the-rtl-sdr/