Phi-3 Weights Released
huggingface.co
huggingface.co
Seems too good to be true, or I'm missing something?
they are updating fixes to the Phi-3 models on a few things very quickly so you will be able test soon but so far in my testing of 4k mini it has some properties that are better than even llama3 70b, namely admitting faults and correcting it once pointed out instead of hiding them again more creatively. Whatever inner intent balancing act it does seems more desirable to me for code things.
[1] https://huggingface.co/microsoft/Phi-3-mini-128k-instruct/di...
and official onnx and gguf (fp16 and Q4) versions
4k will work but "does not support long rope which is use by 128k variant."