granite4.2
ollama pull granite4.2:<size> · sizes supported: 8b
The size listed here runs all six agent surfaces, five consecutive times each: 30 of 30 runs. It carries no settings at all — the second model measured to need none.
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 |
|---|---|---|---|---|---|---|---|---|---|
| 8b | 5.3 GB | 28s | 1:12 | 1:29 | 19s | 45s | 33s | 46s | 1:56 |
Every cell is 5/5, so the table says how long rather than whether.
Slower than granite4.1:8b on the same disk footprint — four times the median — and it is the
same six verdicts at the end. Take 4.1 unless you have a reason to prefer this one.
What it needs that the defaults do not give it
Section titled “What it needs that the defaults do not give it”granite4.2:8b
Section titled “granite4.2:8b”Nothing. Measured at the defaults — temperature 0, top_p 1, a ceiling of twelve unreported calls, the shipped 90-second turn limit — and it clears every surface on them.
It still has an entry in the measurement document, and that is deliberate: an absent entry records no measurement rather than a measurement that found nothing to change, and a model nobody can see was measured is a model nobody can trust.
Where it is thin
Section titled “Where it is thin”Assess re-read 4/5 on a second sitting, having locked at 5/5 on the first. The loss was a report the run never composed, not a wrong answer. Both readings are on record and the cell is counted as passing; a figure that reads steadier than the model is worse than one that says where it is thin.
Method, and where these numbers stop being safe to generalise from:
methodology.md.