Qwen3.6-27B Throughput Slips on 3090

Hot take: most of the impressive TPS claims are for a narrow, favorable setup, not the workload people actually want for coding agents.

• –The gap is real: one-shot generation on short prompts is much faster than repeated, tool-using, long-context agent loops.
• –Long context is expensive even before reasoning starts; once you fill the window, prefill and attention costs dominate and the “headline TPS” becomes much less relevant.
• –Thinking mode is not free. If you want better coding quality, you usually pay for it in latency and lower token throughput.
• –The skepticism about MTP/DFlash-style approaches is reasonable: they can look strong in short demos, but quality and speed both tend to degrade as context and task complexity rise.
• –The likely answer is not “skill issue” alone. It is a mismatch between benchmark framing and real harness behavior, plus the limits of current inference tricks.
• –The practical metric for this use case is not raw TPS, but task completion time at acceptable quality across multi-turn coding jobs.