qwen3.6.md

qwen3.6

ollama pull qwen3.6:27b · one size measured: 27b (17 GB)

Passes all three workflows — and is the clearest evidence on these pages that the machine matters as much as the model.

Results

Size	Disk	Investigate	Operate	Analyze	Score
27b	17 GB	✅ 56s	✅ 108s	✅ 87s	3/3

No small tags are published, same as qwen3.8. For a smaller Qwen see qwen3 or qwen3.5.

It scored 0/3 on the same machine an hour earlier

Every one of the three workflows failed with model-timeout. The ledger showed where the 90-second per-turn ceiling went:

    0.0s  run-started
   42.0s  tool-invoked x4        <- turn 1 took 42s
  132.0s  run-finished           <- turn 2 hit exactly 90s and was cut

Lifting the ceiling temporarily to measure it showed turn 2 needed 91.4 seconds. The model was missing agent mode by 1.4 seconds.

The laptop was in low-power mode. Switched to high power, the same run passed:

	Low power	High power
one turn, small prompt	20.6s	4.6s
throughput	10.3 tok/s	28.9 tok/s
Investigate	❌ model-timeout	✅ 56s
Operate	❌ model-timeout	✅ 108s
Analyze	❌ model-timeout	✅ 87s

No code changed. A per-turn timeout override was drafted and then not written, because the number it was sized against turned out to be an artifact of the power setting.

If a 27B model times out for you, check the power mode before concluding anything about the model. On macOS: pmset -g | grep powermode — 2 is high power. Turn 2 of an agent run is the expensive one, because the tool results from turn 1 are in the prompt by then.