Tag
User reports that Gemma 4 26b outperforms Qwen 3.5/3.6 for language learning and scientific queries, despite being behind in coding tasks, and invites discussion on other non-coding use cases for small MoE models.