There are no viable self-hostable alternatives to GPT-4 or even to GPT3.5.
The “best” self-hostable model is a moving target. As of this writing it’s probably one of Vicuña 13B, Wizard 30B, or maybe Guanaco 65B. I’d like to say that Guanaco is wildly better than Vicuña, what with its 5x larger size. But… that seems very task dependent.
As anecdata: my experience is that none of these is as good as even GPT3.5 for summarization, extraction, sentiment analysis, or assistance with writing code. Figuring out how to run them is painful. The speed at which their unquantized variants run on any hardware I have access to is painful. Sorting through licensing is… also painful.
The “best” self-hostable model is a moving target. As of this writing it’s probably one of Vicuña 13B, Wizard 30B, or maybe Guanaco 65B. I’d like to say that Guanaco is wildly better than Vicuña, what with its 5x larger size. But… that seems very task dependent.
As anecdata: my experience is that none of these is as good as even GPT3.5 for summarization, extraction, sentiment analysis, or assistance with writing code. Figuring out how to run them is painful. The speed at which their unquantized variants run on any hardware I have access to is painful. Sorting through licensing is… also painful.
And again: they’re nowhere close to GPT-4.