Evaluation run for all "good open weight models" with all available quantizations and different GPUs

[zimmski]

See https://www.reddit.com/r/LocalLLaMA/comments/1dlsxab/comment/l9rzjj7/?utm_source=share&utm_medium=web3x&utm_name=web3xcss&utm_term=1&utm_content=share_button

Not sure on how we should do that yet. CPU-only-inference will break us here, and speed-metrics are important as well.