Lately I’ve been submitting the same requests to GPT 4o and Mistral Large running on my laptop or one of my own servers that has 2xA6000. I get about the same q

Lately I’ve been submitting the same requests to GPT 4o and Mistral Large running on my laptop or one of my own servers that has 2xA6000. I get about the same quality from both and probably would not be able to tell which is which in a blind test. There are a couple disadvantages to either of the two local solutions, however. Both of them are slower than 4o. Both of them require a substantial amount of very expensive hardware (which I happed to already have, but still). And on a laptop a single long generation can drain a few percent of the battery. Neural Engine, which would be more energy efficient cannot currently run models this large, and smaller ones do not do as well for the chatbot use case. I do also use smaller LLMs locally, for the more “narrow” tasks.