Qwen3.5-2B that thinks less and answers better: models, data and eval outputs