vibe gpt-oss:20b vs qwen3-coder:30b vs claude sonnet 4.5 . Sorry to report closed models are far better than self hosted. Qwen3-coder isn't bad, but in agent mode it wasn't able to iterate as much. gpt-oss thinks harder and gets better results in the first try, but then also fails to improve much in agent mode. #vibecode #qwen3code #oss-llm