Note that he did not confirm is was mistral-medium. He says that's a retrained llama2-70B model, but hints that it is not the fully trained one. Sounds a bit like damage control but is not a 100% confirmation of the claim.
Only quantized versions of the model were leaked. If you see any unquantized version of it then it's something which was recreated from these, and not the original model. People have also requanted it from GGUF to EXL2 and probably other formats too.
I have gpt4 with copilot at work so i mostly use that, it is pretty good for programming, you see its limitations when you start relying on it a lot on autocompletions, maybe because the code itself is not too coherent and gpt gets confused
Maybe chat tasks are simpler because it writes the whole thing from scratch, though I mostly ask it for limited functionalities, eg write a function that takes x and returns y or use a certain ORM to change a value.
I would suggest you try both gpt 3.5 and the free huggingface minstral 7b version (you can probably run it on pc too, it's not huge) for programming tasks and see for yourself. For general knowledge though, gpt wins hands down over minsral 7b
I've had reasonably good results with deepseek 6.7b
It struggles and hallucinates once you get into more niche languages, but it'll spit out a snake in Python or write file io functions pretty well - I've used it to read/write huge json in a buffer with good results.
I haven't tried gpt4, but compared to 3.5 it's pretty great
