Deepseek 67b is amazing, and in at least 1 usecase it seems better than ChatGPT 4
Just wanted to toss this out there since I don't see a lot of folks talking about it, but Deepseek 67b Chat has become one of my favorite general purpose models.
This is also the first model I've run into that has clearly beaten out, for me at least, ChatGPT 4 in a use case: Excel and VBA.
My wife has been working on a very complex excel worksheet, with a lot of automation involved in it, and we found that most open source models are AWFUL at answering excel questions; at best, local models were giving barely acceptable answers that only partially worked.
In fact, ChatGPT 4 was only giving just acceptable answers... it at least put us on the right track for some stuff, but ultimately wasn't 't doing the trick entirely.
Deepseek 67b Chat was the first model to give actually good answers for excel and vba stuff. Not great, not perfect, but better than any AI we'd tried yet.
In general, I've found this is the case with open source models. I've yet to find a local model as good as ChatGPT 4 in everything, but as time goes we're identifying models good at something specific; usually at best they are almost as good as ChatGPT 4, but this is a unique case where it feels actually better, in my opinion.
So I wanted to throw that out there. We're running the q8 of it, but I imagine down to q4 is still really good.
Edit: Updated post to point out that this is Deepseek 67b Chat, not the base. I haven't put any time into testing the base, so unsure of its quality.
Edit 2: I'm using Oobabooga's Starchat preset (0.2 temp) and the Deepseek instruct template (automatically loads with the model in Ooba).