Mixtral 8x22B Instruct is out. It significantly outperforms existing open models, and only uses 39B active parameters (making it significantly faster than 70B models during inference). 1/n

→ View original post on X — @guillaumelample,