8B models are pretty fast even on something like a 3060 depending on deployment method (for example Q4 on Ollama).