nemotron3
ollama pull nemotron3:<size> · sizes supported: 33b
The size listed here runs all six agent surfaces, five consecutive times each: 30 of 30 runs. The largest model in the product by a wide margin, and the only one that ended a run by asking its user a question.
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 |
|---|---|---|---|---|---|---|---|---|---|
| nemotron3:33b | 27 GB | 12s | 49s | 20s | 18s | 29s | 14s | 21s | 1:59 |
Optimize is the cell that took twenty-eight runs across eight sittings to close, and it is the one to read the note below about.
What it needs that the defaults do not give it
Section titled “What it needs that the defaults do not give it”nemotron3:33b
Section titled “nemotron3:33b”a retry when it stops having read nothing. Twice it ended a run by asking for the data
instead of reading it — the clearer of the two ten seconds in, zero tools invoked, closing with
“Could you please share the exact SQL statement you’re running for the employee listing?” while
holding inspect_schema and inspect_plan. A run has no user on the other end, so the question
is a stop. The drive now names the instrument once and gives the turn back.
Free to grant: compose_report is one of the tools the drive counts, so a run reaching that
point with nothing called has already earned no-report, and the turn is spent on a run that
has lost.
worked refusal examples. Three of its optimization losses are one refusal repeated until the
run runs out of time — claims: expected array, the same sentence each time, seven times in one
run. The example is built from the run’s own ledger, so a model that copies it verbatim gets a
call that is accepted.
Where these figures are softest
Section titled “Where these figures are softest”The optimization cell is locked but marginal, and every sweep since has said so: 5/5, 5/5, 4/5, 4/5, and 3/5 on a rested machine. Every loss has one signature — the turn after the last tool call running long, at 90.0 s, 106 s, 172.7 s, 244.6 s and 430.8 s against a 90-second turn, where the passing runs compose in 21 to 100.
A 150-second turn was measured against it directly (#500): five runs read 3/5 again, but one of them composed at 114 s — a run the shipped limit would have killed. So the limit is part of the answer and not all of it; what the runs actually show is a report turn whose length varies by a factor of twenty on identical input. Assess shows the same shape at 3/5 on the same sweep.
Recorded rather than smoothed: the pass/fail figures above are the measurement that locked the cell, and these are what a later reader gets when they run it again.
Method, and where these numbers stop being safe to generalise from:
methodology.md.