All of the LLaMA variants like Vicuna are finetunes, hence are tainted by base LLaMA. Stability AI's model is not good, and generally the LLaMA clones are not

All of the LLaMA variants like Vicuna are finetunes, hence are tainted by base LLaMA. Stability AI's model is not good, and generally the LLaMA clones are not ready yet. There is a really cool model that actually outperforms LLaMA: Falcon 7B and Falcon 40B. But it is extremely new, and not even supported by llama.cpp (the backend many projects use) yet, though it is on the June roadmap. You would have to run the 4 bit variant on an Nvidia GPU, and you might have to finetune it for certain llm frameworks.