Qwen3.5 uses Gated Delta Networks which is essentially Mamba 2 + Delta Rule. It’s quite hardware efficient.

> Is it? In what ways?

Just the reinforcement learning for reasoning, and then tool use for agents, could be its own topic.