Qwen3.5-27B Slows Despite Dense Design

30 tok/s is not the real headline here; the 30-95 second TTFT is.

• –Qwen3.5-27B is dense, so every generated token uses all 27B parameters. MoE models with far fewer active parameters can be dramatically faster at the same nominal size.
• –Qwen’s docs say Qwen3.5 defaults to thinking mode, which can add hidden reasoning work before the visible answer unless the provider disables it.
• –OpenRouter’s provider table is routed serving, not a clean single-GPU benchmark, so batching, queue depth, prompt length, and CPU offload can swing throughput a lot.
• –TTFT means time to first token, so those long waits include prompt processing and queueing as well as model compute.
• –On consumer cards, a 27B model often does not fit comfortably at useful precision, so once layers spill out of VRAM, token speed drops fast.