Released Mixtral 8x7B, GPT3.5 or Llama 2 70B perf with 12B inference speed & cost