reasoningopinionneutralPost-training recipes differ between Qwen3.5 and Gemma4 models, producing different reasoning behaviorsLewis Tunstall30 Jul 2026https://x.com/_lewtun/status/2082832646745317422