qwen3.5
ollama pull qwen3.5:<size> · sizes supported: 4b, 9b, 27b
Every size listed here runs all six agent surfaces, five consecutive times each: 30 of 30 runs.
What it does, and how long it takes
Section titled “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 |
|---|---|---|---|---|---|---|---|---|---|
| 4b | 3.4 GB | 5s | 41s | 21s | 7s | 34s | 1s | 11s | 2:47 |
| 9b | 5.8 GB | 31s | 31s | 52s | 21s | 26s | 1:38 | 33s | 1:38 |
| 27b | 17 GB | 78s | 91s | 149s | 41s | 66s | 48s | 78s | 2:29 |
Every cell is 5/5, so the table says how long rather than whether.
The 4b is three times the 9b’s speed at two thirds of its disk, which is not the direction size usually runs. Its tail is the price: 2:47 against the 9b’s 1:38.
The 27b is the one that asked for nothing. It cleared all six surfaces on its first attempt at
the compiled defaults, the whole model in 52 minutes, with not one setting spent — the only size in
this family whose row in measured-profiles.json carries no departure from the defaults at all. It
is also the slowest of the three, and both facts are worth stating together: what a model costs in
seconds and what it costs in configuration are separate questions, and this one answers them in
opposite directions.
Its optimize cell is the figure to notice. Three other models measured the same week could not open that surface at any setting; this one took it 5/5 unaided.
What it needs that the defaults do not give it
Section titled “What it needs that the defaults do not give it”qwen3.5:4b
Section titled “qwen3.5:4b”no reasoning on the plan turn. Its plan turns timed out on every run with an empty ledger —
nothing called, nothing said. Suppressing the reasoning request finished the turns and moved the
cell 0/5 to 1/5. It reaches the OpenAI-compatible adapter only, so it is a no-op on gemini.
The other four losses were not the model’s, and this is the only cell in the product closed by
fixing dbportal’s own wording. It had been writing correct refusals — naming the views whose columns
the inventory cannot derive, then asking the one question that would unblock it — and opening every
one with NO STATEMENT AT ALL:, which is the phrase the planning rule itself put in front of the
marker it was teaching. The rule now says it once, and the cell read 5/5 on the first pass after.
Where it is thin
Section titled “Where it is thin”The 4b’s data-analysis re-read 3/5 on a second sitting, having locked at 5/5 on the first. The cell has locked repeatedly and is counted as passing; both readings are on record, because a figure that reads steadier than the model is worse than one that says where it is thin. If you drive this size on data-analysis, expect it to need a second attempt more often than the table’s 11s suggests.
qwen3.5:9b
Section titled “qwen3.5:9b”a 150-second turn limit. Five surfaces clear the shipped 90 seconds comfortably; its plan turn lands at 92 to 94. It was the first model to need it and is now one of seven that carry it; the limit stays 90 for the other twenty-one.
empty-turn retry. Its optimization runs stopped answering after being corrected, leaving no text behind.
Method, and where these numbers stop being safe to generalise from:
methodology.md.