granite4.1
ollama pull granite4.1:<size> · sizes supported: 8b, 30b
Every size listed here runs all six agent surfaces, five consecutive times each: 30 of 30 runs. The fastest family measured, and both supported sizes clear every surface.
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 | 6s | 10s | 7s | 11s | 16s | 5s | 11s | 22s |
| 30b | 17 GB | 16s | 42s | 21s | 31s | 26s | 26s | 26s | 52s |
Every cell is 5/5, so the table says how long rather than whether.
What it needs that the defaults do not give it
Section titled “What it needs that the defaults do not give it”granite4.1:8b
Section titled “granite4.1:8b”worked refusal examples. A refused call comes back with a correctly-shaped one built from this run’s own ledger. It knows what to record and cannot always build the object carrying it.
empty-turn retry. Its optimization runs were corrected once and then returned nothing at all — no call, no text. An empty turn reads as a model that stopped; asked again, the cell went 3/5 to 5/5.
granite4.1:30b
Section titled “granite4.1:30b”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.