ollama pull granite4.1:<size> · sizes measured: 3b, 8b, 30b
The fastest family measured by a wide margin — and the one where speed does not carry over to the analysis question.
| Size | Disk | Investigate | Operate | Analyze | Score |
|---|---|---|---|---|---|
| 3b | 2.1 GB | ✅ 2.8s | ✅ 1.3s | ❌ no-report |
2/3 |
| 8b | 5.3 GB | ✅ 4.0s | ✅ 2.9s | ❌ empty-evidence |
2/3 |
| 30b | 17 GB | ✅ 6.7s | ✅ 23.7s | ✅ 19.9s | 3/3 |
granite4.1:3b finishing Operate in 1.3 seconds is the fastest run recorded on
any model here. For questions that are only about the engine's own state, at 2.1 GB,
that is hard to beat.
3b — no-report. It reads, then writes its findings as prose instead of calling
compose_report. The runtime reminds such a run once; this one narrates again.
8b — empty-evidence. More interesting: it composes a report, but the claims
cite nothing the run established. This is the citation contract doing its job rather
than a defect. A report whose claims are not backed by a reading this run took is
refused, and that refusal is the feature — it is what stops a confident summary of a
database nobody looked at.
Neither is fixable by a nudge. empty-evidence in particular should not be: loosening
it would let exactly the failure the contract exists to prevent through.
Use granite4.1:30b if you want the whole workflow set. Use granite4.1:3b
deliberately, for operational questions only, where its speed is remarkable.