That depends on your goal and available memory. I generally use Mixtral for text, LlaVA for multimodal stuff, and Copilot in VSCode.