ministral-3

ollama pull ministral-3:<size> · sizes supported: 8b, 14b

Every size listed here runs all six agent surfaces, five consecutive times each: 30 of 30 runs. Mistral’s first entry on this list, and both supported sizes cleared every surface on their first attempt with nothing configured.

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	6.0 GB	16s	22s	4s	14s	13s	5s	13s	26s
14b	9.1 GB	19s	27s	34s	13s	15s	8s	16s	56s

Every cell is 5/5, so the table says how long rather than whether.

The 8b is the faster of the two everywhere except operate, and the gap is widest on assess — 4 seconds against 34. Between two sizes that both clear everything, the smaller one is the one to take; the 14b is here because a family that clears six surfaces at one size is worth measuring at the next, not because it does anything the 8b cannot.

What it needs that the defaults do not give it

ministral-3:8b

Nothing. Measured at the defaults — temperature 0, top_p 1, a ceiling of twelve unreported calls — and it clears every surface on them.

ministral-3:14b

Nothing. Measured at the defaults, and it clears every surface on them.

The 3b is not supported

ministral-3:3b locks three of the six — analyze, assess and plan — and does not lock investigate, operate or optimize. It is not that it cannot answer them: across nine attempts at those three cells it passed 20 of 45 runs, reading 4/5 on investigation once. The losses are one shape, no-report: it reads the database, and then finishes the run without filing what it read. Five consecutive is the bar precisely because a model that answers four times in five is not one to put in front of a user, so the 3b is measured, recorded here, and not listed as supported.

---

Method, and where these numbers stop being safe to generalise from: methodology.md.