Good question! I am collecting human data on how quantization affects outputs. See here for more information: ggml-org/llama.cpp#5962
In the meantime, use the largest that fully fits in your GPU. If you can comfortably fit Q4_K_S, try using a model with more parameters.
See the wiki upstream: https://github.com/ggerganov/llama.cpp/wiki/Feature-matrix
Since we're dealing with simpler AV1 encoding, that does mean we'll be eskewing aomenc-av1, since it requires 2-pass encoding to be able to take advantage of it, and I use it externally since I have access to a special build.
Now, let's get on with the simple guide.
You'll first need to be reasonably competent with command line builds or use a recent up to date ffmpeg GUI with support for SVT-AV1.
As such, I would recommend getting a master git ffmpeg build for the operating system of your choice right here:
