It would be very surprising. Mistral AI's github page has more information on their sliding window attention method to achieve this performance: https://github

It would be very surprising. Mistral AI's github page has more information on their sliding window attention method to achieve this performance: https://github.com/mistralai/mistral-src https://github.com/mistralai/mistral-src If Mistral 7b lives up to the claims, I expect these techniques will make their way into llama.cpp. But I would be surprised if the required updates were quick or easy.