qwen3

ollama pull qwen3:<size> · sizes supported: 8b, 14b, 4b

Every size listed here runs all six agent surfaces, five consecutive times each: 30 of 30 runs. Three sizes supported, and the smallest is not the fastest — size buys memory rather than speed on this workload.

What it does, and how long it takes

Seconds are the median of the runs that passed, per surface.

Size	Disk	Investigate	Optimize	Assess	Operate	Analyze	Plan	Median	Slowest
8b	5.2 GB	10s	1:18	46s	36s	46s	26s	36s	1:48
14b	9.3 GB	15s	3:10	1:22	46s	31s	21s	41s	5:03
4b	2.5 GB	41s	2:04	1:13	1:12	1:43	46s	1:12	2:40

Every cell is 5/5, so the table says how long rather than whether.

What it needs that the defaults do not give it

qwen3:8b

sampling on optimization only. The only sampled cell in the product. At temperature 0 it opened with the wrong tool ten times out of ten; at 0.8 it opened with the right one and won. Its other five surfaces are deterministic.

qwen3:14b

one extra plan turn. The model this setting was written for: five surfaces locked and plan lost four times in five, always for the same missing sentence.

qwen3:4b

Nothing. Measured at the defaults — temperature 0, top_p 1, a ceiling of twelve unreported calls — and it clears every surface on them.

---

Method, and where these numbers stop being safe to generalise from: methodology.md.