Qwen 3.5 tends to overthink, which ironically helps it better infer intent Gemma 4 is the opposite, you have to spell everything out (probably guardrails) Had Qwen rewrite my prompts, then used those on both models Gemma’s performance jumped noticeably

→ View original post on X — @theahmadosman