>>82702421That was understandable a while back when Llama was still trying to make up ground in the open source community after GPT-3.5 was deemed to never go open-source and a lot of companies followed after OpenAI and trying to set up moats around their models. But after this summer with the flurry of releases after Llama 3, the gap has closed considerably between local and cloud looking at ChatArena Hard queries benchmark. Even if you can't run Mistral Large 2 (MIstral-Large-2407), Gemma 2 27B is good enough to equal Claude 3 Opus levels of performance if not better especially with specific finetunes. A difference of 22% winrate is considerable but it's not a crippling difference. I remember when models based on Llama 1 was way behind where Claude 1 and the initial GPT-4 release with hundreds of elo points difference
https://lmsys.org/blog/2023-05-25-leaderboard/. The difference using the same metrics is between Gemma 2 and Claude 3.5 is now around 50 points. I don't think anyone would do a comparison of what you can get between local and cloud now and decide to send dickpics to get access to the latter now. The main issues with local LLMs now is to match GPT4-o with multimodality and that is still a huge gap but text should be enough for RP purposes.
